Water surface target edge detection algorithm based on evidence fusion driving

Abstract

The invention discloses a water surface target edge detection algorithm based on evidence fusion driving, and the algorithm comprises the following steps: collecting a water surface target sample image through a visual sensor, carrying out the graying of the water surface target sample image, and obtaining a gray image Iray (x, y); suppressing high-frequency noise in the grayscale image by using Gaussian smoothing filtering; constructing an edge feature extraction operator to extract a gradient magnitude M (x, y) and a direction theta (x, y) of the image; a probabilistic non-maximum suppression mechanism is constructed, so that edge judgment is credible; performing bilateral threshold on the gradient amplitude, and dividing edges into a strong edge, a weak edge and a non-edge; determining the probability of dividing weak edges into edges under different evidence sources from three aspects of gradient magnitude, gradient direction and neighborhood connectivity; s2, calculating a joint confidence coefficient m'e (x, y), and determining a credible weak edge as an edge according to the joint confidence coefficient, thereby obtaining the final edge information of the image; according to the invention, the detection precision of the water surface target contour in a complex environment can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to a water surface target edge detection algorithm based on evidence fusion driving. BACKGROUND

[0002] With the continuous development of intelligent perception and autonomous control technology, water unmanned systems are widely used in military reconnaissance, water security, hydrological monitoring, water rescue and intelligent transportation and other fields. In water unmanned perception systems, edge information is the basic data for target recognition, segmentation and subsequent high-level perception processing. High-robustness and high-precision edge detection algorithm has become an important research topic in the field of water surface visual perception.

[0003] In the water surface scene, due to the influence of natural complex factors such as wave fluctuation, reflection interference, and uneven illumination, the traditional edge detection method has obvious limitations in edge positioning and noise resistance. As a classic edge detection algorithm, Canny algorithm improves the edge positioning accuracy and continuity by non-maximum suppression (NMS) and double threshold hysteresis connection. However, its non-maximum suppression strategy is heavily dependent on the judgment of local gradient extreme value, and lacks the introduction of regional context information, so the edge detection algorithm is easily disturbed by noise; in addition, when the gradient amplitude fluctuates, the continuous edge is easily "cut off" in the local maximum detection, resulting in serious edge breakage. In the double threshold edge connection process, the basis for judging whether the weak edge is an edge: if there is a strong edge in its neighborhood, the weak edge is judged as an edge, and this decision method is easily affected by noise. Therefore, it is urgent to design an edge detection method combined with regional context information to effectively improve the detection accuracy, continuity and robustness of water surface target contour in complex environment. SUMMARY

[0004] A water surface target edge detection algorithm based on evidence fusion driving includes the following steps:

[0005] Step A: Collect the water surface target sample image through the visual sensor, and gray it to obtain the gray image I gray (x,y);

[0006] Step B: Use Gaussian smoothing filter to suppress high-frequency noise in the gray image, making the edge detection more stable;

[0007] Step C: Construct an edge feature extraction operator to extract the gradient amplitude M(x,y) and direction θ(x,y) of the image;

[0008] Step D: Construct a "probabilistic non-maximum suppression mechanism" to make the edge decision credible, and record the gradient magnitude at pixel point (x, y) that may be an edge as M1(x, y);

[0009] Step E: Perform bilateral thresholding on the gradient magnitude M1(x, y) at pixel point (x, y) to further divide the pixel point that may be an edge into a strong edge, a weak edge and a non-edge;

[0010] Step F: Determine the probability of dividing the weak edge into an edge under different evidence sources from three aspects of gradient magnitude, gradient direction and neighborhood connectivity;

[0011] Step G: According to the probability of dividing the weak edge into an edge under different evidence sources in Step F, construct a decision fusion method S2 to calculate the joint confidence m'; e (x, y) according to the joint confidence, determine the credible weak edge as an edge, and thus obtain the final edge information of the image;

[0012] The method for suppressing high-frequency noise in the gray image by using Gaussian smoothing filtering in Step B is as follows:

[0013] Firstly, a Gaussian function S(x, y) is constructed, and Gaussian smoothing filtering is used to suppress the high-frequency noise in the gray image I gray (x, y), so that the edge detection is more stable, and the specific method is as follows:

[0014]

[0015] wherein σ is a standard deviation, which determines the width of the filter. Then, a 3x3 Gaussian template S'(x, y) is constructed by using the Gaussian function S(x, y) with the current pixel point as the coordinate origin. Finally, the gray image I gray (x, y) is convolved by using the Gaussian template S'(x, y) to obtain the smoothed image I1(x, y), and the specific method is as follows:

[0016] I1(x, y) = I gray (x, y) * S'(x, y)

[0017] The method for constructing an edge feature extraction operator in Step C to extract the gradient magnitude M(x, y) and direction θ(x, y) of the image is as follows:

[0018] Step C1: Perform a difference operation on the smoothed image I1(x, y) to extract the gradient image G x (x, y) in the x direction and the gradient image G y (x, y) in the y direction, and the specific method is as follows:

[0019]

[0020] Step C2: According to the G x (x,y) and G y (x,y), the gradient magnitude M(x,y) and the gradient direction θ(x,y) of the image are calculated, specifically as follows:

[0021]

[0022] The step D builds a "probabilistic non-maximum suppression mechanism" in the method, so that the edge decision can be trusted, which is:

[0023] Step D1: Based on the similarity of the gradient direction and the maximum gradient in the neighborhood, the category confidence m1 of the pixel point is determined, specifically as follows:

[0024] Firstly, the gradient direction θ(x,y) of each point of the image calculated in step C2 is divided into an interval every 45 degrees, and the gradient direction of each point of the intervalized image is denoted as θ'(x,y);

[0025] Then, for any pixel point (x,y) in the image, the gradient magnitude matrix M(x,y) of the neighboring pixels corresponding to the gradient direction θ'(x,y) of the pixel point is obtained, and the maximum gradient magnitude in M(x,y) is denoted as M max (x,y);

[0026] Secondly, according to the proximity degree of M(x,y) and M max (x,y), the probability m 1e (x,y) that the point is an edge is calculated, specifically as follows:

[0027] d1(x,y) = M(x,y) - M max (x,y)

[0028]

[0029] Wherein, d1(x,y) represents the distance between M(x,y) and M max (x,y), i.e. the proximity degree; m 1e (x,y) represents the probability that the pixel point is an edge pixel point; m 1n (x,y) represents the probability that the pixel point is a non-edge pixel point; and α represents a parameter adjustment factor.

[0030] Step D2: Based on the neighborhood edge continuity, the category confidence m2 of the pixel point is built, specifically as follows:

[0031] Firstly, for any pixel point (x,y) in the image, the m 1e(x+i, y+j) > thre1, f1() = 1, otherwise 0.

[0032]

[0033] where f1() is an indicator function, when m 1e (x+i, y+j) > thre1, f1() = 1, otherwise 0.

[0034] Then, according to each pixel number N1(x, y), the cascade probability m 2e (x, y) of the point being an edge is calculated

[0035]

[0036] where m 2e (x, y) represents the probability of the pixel point being an edge; m 2n (x, y) represents the probability of the pixel point not being an edge.

[0037] Step D3: According to the probability of each pixel point being judged as an edge in steps D1 and D2, a credibility edge decision fusion method S1 is constructed to obtain the fusion confidence m e (x, y) of each pixel point being judged as an edge, as follows:

[0038]

[0039] where m e (x, y) represents the fusion confidence of the pixel point being judged as an edge, m n (x, y) represents the fusion confidence of the pixel point not being an edge.

[0040] Step D4: According to the fusion confidence in step D3, in combination with a confidence threshold thre2, the pixel points not being edges in the image are removed, the gradient amplitude M1(x, y) of the pixel points possibly being edges is recorded, the "probabilistic non-maximum suppression" is realized, and the image edges are refined, as follows:

[0041]

[0042] The method for dividing the edges into strong edges, weak edges and non-edges by performing bilateral thresholding on the gradient amplitude M1(x, y) of step D in step E is as follows:

[0043]

[0044] where thre H and thre LM2(x, y) = 255 represents that the pixel point is a strong edge, M2(x, y) = 75 represents that the pixel point is a weak edge, and M2(x, y) = 0 represents that the pixel point is not an edge;

[0045] The method for determining the probability of classifying the weak edge in step E as an edge from three evidence sources of the gradient amplitude, the gradient direction and the neighborhood connectivity in step F is as follows:

[0046] Step F1: constructing an evidence source according to the gradient amplitude; according to step E, the gradient amplitude of the weak edge is between threshold thre H and thre L In addition, according to the greater the gradient amplitude of the current pixel point, the more likely the pixel point is an edge point, the probability of classifying the weak edge pixel point as an edge in the gradient amplitude evidence source is as follows:

[0047]

[0048] Wherein, m 3e (x, y) represents the probability of classifying the weak edge pixel point (x, y) as an edge; and m 3n (x, y) represents the probability of classifying the weak edge pixel point (x, y) as a non-edge.

[0049] Step F2: constructing an evidence source according to the gradient direction; if the gradient direction of the weak edge pixel point is similar to the gradient direction of the strong edge pixel point in the neighborhood, the probability of the weak edge pixel point being an edge is higher, and the specific steps are as follows:

[0050] First, count the number count4(x, y) of the strong edge pixel points in the neighborhood of the weak edge pixel point, and the specific steps are as follows:

[0051]

[0052] Wherein, f2(·) is an indicator function, which is 1 when M2(x+x i ,y+y i ) = 255, and 0 otherwise;

[0053] Then, the similarity between the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in the neighborhood is calculated, and the specific steps are as follows:

[0054]

[0055] Wherein, the higher the similarity between the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in the neighborhood, the smaller sim4(x, y) is;

[0056] Finally, the probability of the weak edge pixel point being judged as an edge under the gradient direction evidence source is determined according to count4(x, y) and sim4(x, y) as follows:

[0057]

[0058] wherein, m 4e (x, y) represents the probability of the weak edge pixel point (x, y) being judged as an edge under the gradient direction evidence source; m 4n (x, y) represents the probability of the weak edge pixel point (x, y) being judged as a non-edge under the gradient direction evidence source. represents the evidence strength, that is, when there are strong edges in the neighborhood of the weak edge pixel point, the greater the number of strong edges, the greater the probability of the weak edge pixel point being judged as a strong edge. represents the gradient direction similarity support term, that is, the higher the similarity between the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in its neighborhood, the smaller sim4(x, y) is, the greater; when there are no strong edges in the neighborhood of the weak edge pixel point, the probability of the weak edge pixel point being judged as a strong edge is 0.

[0059] Step F3: constructing the evidence source according to the neighborhood connectivity. According to the edge continuity, if the similarity between the gradient amplitude of the weak edge pixel point and the gradient amplitude of the strong edge pixel point in its neighborhood is high, the probability of the weak edge pixel point being an edge is higher, and the specific steps are as follows:

[0060] Firstly, the similarity between the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood is calculated, which is as follows:

[0061]

[0062] wherein, the higher the similarity between the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood, the smaller sim5(x, y) is.

[0063] Then, the sum Nor(x, y) of the gradient amplitudes calculated in step D4 corresponding to each strong edge pixel point in the neighborhood is calculated, which is as follows:

[0064]

[0065] Finally, the probability of the weak edge pixel point being judged as an edge under the neighborhood connectivity evidence source is determined according to count4(x, y), Nor(x, y) and sim5(x, y) as follows:

[0066]

[0067] wherein, m 5e(x, y) represents the probability that the weak edge pixel point (x, y) is judged as an edge under the neighborhood connectivity evidence source; m 5n (x, y) represents the probability that the weak edge pixel point (x, y) is judged as a non-edge under the neighborhood connectivity evidence source; represents the evidence strength, that is, when there is a strong edge in the neighborhood of the weak edge pixel point, the greater the number of strong edges, the greater the probability that the weak edge pixel point is judged as a strong edge; represents the gradient amplitude similarity support term, that is, the higher the similarity between the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood, the smaller sim5(x, y) is, the greater;

[0068] The step G is to construct a decision fusion method S2 according to the probability that the weak edge is divided into an edge under different evidence sources in the step F, and to calculate a joint confidence m' e (x, y), and to determine the credible weak edge as an edge according to the joint confidence, so as to obtain the final edge information of the image, as follows:

[0069] Step G1: according to the probability that each pixel point is judged as an edge in steps F1, F2 and F3, a decision fusion method S2 is constructed to obtain the joint confidence m' e (x, y) of the weak edge pixel point, as follows:

[0070] m' e (x, y) = m 3e × m 4e × m 5e

[0071] Step G2: according to the joint confidence m' e (x, y) in step G1, it is determined whether the weak edge pixel point is an edge. For the weak edge pixel point (x, y), if the joint confidence m' e (x, y) corresponding to the weak edge pixel point is greater than a threshold pre, the weak edge pixel point is determined as an edge, otherwise as a non-edge, as follows:

[0072]

[0073] wherein M2(x, y) is the edge of the image. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is a schematic diagram of a part of a water surface target sample image.

[0075] Figure 2 is a flowchart of a water surface target edge detection algorithm based on evidence fusion driving.

[0076] Figure 3Part result graph of water surface target edge detection algorithm based on evidence fusion driving. DETAILED DESCRIPTION

[0077] The application will be further described in connection with the embodiments.

[0078] As Figures 1-3 shown: the application discloses a water surface target edge detection algorithm based on evidence fusion driving, which combines Figure 2 , and specifically includes the following steps:

[0079] Step A: collect water surface target sample images through a visual sensor, as Figure 1 shown, including island, ship, beach and other scenes, and gray-scale the collected images to obtain a gray image I gray (x,y).

[0080] Step B: use Gaussian smoothing filtering to suppress high-frequency noise in the gray image I gray (x,y), so that the edge detection is more stable, and the specific method is as follows:

[0081] Step B1: construct a Gaussian function S(x,y), specifically as follows:

[0082]

[0083] Wherein, σ is the standard deviation, which determines the width of the filter.

[0084] Step B2: use the Gaussian function S(x,y) to construct a 3*3 Gaussian template S'(x,y) with the current pixel point as the coordinate origin;

[0085] Step B3: use the Gaussian template S'(x,y) to convolve the gray image I gray (x,y) to obtain a smoothed image I1(x,y), specifically as follows:

[0086] I1(x,y) = I gray (x,y)*S'(x,y)

[0087] Step C: construct an edge feature extraction operator to extract the gradient amplitude M(x,y) and direction θ(x,y) of the image, and the specific method is as follows:

[0088] Step C1: perform difference operation on the smoothed image I1(x,y) to extract the gradient image G x (x,y) in x direction and the gradient image G y (x,y) in y direction, specifically as follows:

[0089]

[0090] Step C2: using gradient image G x (x,y) and G y (x,y), the gradient magnitude M(x,y) and the gradient direction θ(x,y) of the image are calculated, as follows:

[0091]

[0092] wherein the gradient magnitude M(x,y) represents the edge intensity of the pixel point (x,y), and the gradient direction θ(x,y) represents the edge direction of the pixel point (x,y).

[0093] Step D: constructing a "probabilistic non-maximum suppression mechanism" to make the edge determination credible, as follows:

[0094] Step D1: determining the category confidence m1 of the pixel point based on the similarity of the gradient direction and the maximum gradient of the neighborhood, as follows:

[0095] First, the gradient direction θ(x,y) of each point of the image calculated in step C2 is divided into an interval every 45 degrees, and the gradient direction of each point of the intervalized image is denoted as θ'(x,y);

[0096] Then, for any pixel point (x,y) in the image, the gradient magnitude matrix M(x,y) of the neighboring pixels in the gradient direction θ'(x,y) corresponding to the pixel point is obtained, and the maximum gradient magnitude in M(x,y) is denoted as M max (x,y);

[0097] Secondly, the probability m max (x,y) that the point is an edge is calculated according to the degree of similarity between M(x,y) and M 1e (x,y), as follows:

[0098] d1(x,y) = M(x,y) - M max (x,y)

[0099]

[0100] wherein d1(x,y) represents the distance between M(x,y) and M max (x,y), i.e. the degree of similarity; m 1e (x,y) represents the probability that the pixel point is an edge pixel point; m 1n (x,y) represents the probability that the pixel point is a non-edge pixel point; and α represents a parameter adjustment factor, wherein α = 5 in the algorithm;

[0101] Step D2: constructing the category confidence m2 of the pixel point based on the continuity of the neighborhood edge, as follows:

[0102] First, for any pixel point (x, y) in the image, the number N1(x, y) of pixels in the 3x3 neighborhood of the pixel point (x, y) whose m 1e (x±1,y±1) is greater than a confidence threshold thre1 is counted, and the specific process is as follows:

[0103]

[0104] wherein f1(·) is an indicator function, and when m 1e (x+i,y+j) > thre1, f1(·) is 1, otherwise 0.

[0105] Then, according to the N1(x, y) of each pixel point, the cascade probability m 2e (x, y) that the pixel point is an edge is calculated, and the specific process is as follows:

[0106]

[0107] wherein m 2e (x, y) represents the probability that the pixel point is an edge; and m 2n (x, y) represents the probability that the pixel point is not an edge.

[0108] Step D3: According to the probability that each pixel point is judged as an edge in steps D1 and D2, a credibility edge decision fusion method S1 is constructed to obtain the fusion confidence m e (x, y) that each pixel point is judged as an edge, and the specific process is as follows:

[0109]

[0110] wherein m e (x, y) represents the fusion confidence that the pixel point is judged as an edge, and m n (x, y) represents the fusion confidence that the pixel point is not an edge.

[0111] Step D4: According to the fusion confidence in step D3, in combination with a confidence threshold thre2, the pixel points that are not edges in the image are removed, the gradient amplitude M1(x, y) of the pixel points that are possibly edges is recorded, the “probabilistic non-maximum suppression” is realized, and the image edge is refined, and the specific process is as follows:

[0112]

[0113] Step E: The gradient amplitude M1(x, y) is subjected to a bilateral threshold, the edge is divided into a strong edge, a weak edge and a non-edge, and the specific process is as follows:

[0114]

[0115] M2(x, y) = 255 represents that the pixel point is a strong edge, M2(x, y) = 75 represents that the pixel point is a weak edge, and M2(x, y) = 0 represents that the pixel point is a non-edge, thre H thre L represents a gradient amplitude threshold value;

[0116] Step F: constructing evidence sources from three aspects of gradient amplitude, gradient direction and neighborhood connectivity, determining the probability of dividing the weak edge into an edge under different evidence sources, and the specific steps are as follows:

[0117] Step F1: the greater the gradient amplitude of the weak edge pixel point, the more obvious the change of the image at the point, and the greater the probability that the point is an edge point, and the specific steps are as follows:

[0118] According to the gradient amplitude, the evidence source is constructed; it is known from step E that the gradient amplitude of the weak edge is located between the threshold values thre H thre L In addition, according to the greater the gradient amplitude of the current pixel point, the more likely the pixel point becomes an edge point, according to the weak edge pixel point being judged as an edge under the gradient amplitude evidence source, the probability is as follows:

[0119]

[0120] Wherein, m 3e (x, y) represents the probability of the weak edge pixel point (x, y) being judged as an edge; m 3n (x, y) represents the probability of the weak edge pixel point (x, y) being judged as a non-edge.

[0121] Step F2: constructing an evidence source according to the gradient direction. If the gradient direction of the weak edge pixel point is similar to the gradient direction of the strong edge pixel point in its neighborhood, the probability that the weak edge pixel point is an edge is higher, and the specific steps are as follows:

[0122] First, the number of strong edge pixel points count4(x, y) in the neighborhood of the weak edge pixel point is counted, and the specific steps are as follows:

[0123]

[0124] Wherein, f2(·) is an indicator function, which is 1 when M2(x+x i ,y+y i ) = 255, otherwise 0;

[0125] Then, the similarity between the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in its neighborhood is calculated, and the specific steps are as follows:

[0126]

[0127] wherein the higher the similarity between the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in its neighborhood, the smaller sim4(x, y) is;

[0128] Finally, according to count4(x, y) and sim4(x, y), the probability of the weak edge pixel point being judged as an edge under the gradient direction evidence source is determined as follows:

[0129]

[0130] wherein m 4e (x, y) represents the probability of the weak edge pixel point (x, y) being judged as an edge under the gradient direction evidence source; m 4n (x, y) represents the probability of the weak edge pixel point (x, y) being judged as a non-edge under the gradient direction evidence source;

[0131] Step F3: constructing the evidence source according to neighborhood connectivity. According to the edge continuity, if the similarity between the gradient magnitude of the weak edge pixel point and the gradient magnitude of each strong edge pixel point in its neighborhood is high, the probability of the weak edge pixel point being an edge is higher, and the specific steps are as follows:

[0132] Firstly, the similarity between the gradient magnitude of the weak edge pixel point and the gradient magnitude of each strong edge pixel point in its neighborhood is calculated, which is specifically as follows:

[0133]

[0134] wherein the higher the similarity between the gradient magnitude of the weak edge pixel point and the gradient magnitude of each strong edge pixel point in its neighborhood, the smaller sim5(x, y) is;

[0135] Then, the sum Nor(x, y) of the gradient magnitudes calculated in step D4 corresponding to each strong edge pixel point in the neighborhood is calculated, which is specifically as follows:

[0136]

[0137] Finally, according to count4(x, y), Nor(x, y) and sim5(x, y), the probability of the weak edge pixel point being judged as an edge under the neighborhood connectivity evidence source is determined as follows:

[0138]

[0139] wherein m 5e (x, y) represents the probability of the weak edge pixel point (x, y) being judged as an edge under the neighborhood connectivity evidence source; m 5n (x, y) represents the probability of the weak edge pixel point (x, y) being judged as a non-edge under the neighborhood connectivity evidence source;

[0140] Step G: According to the probability of weak edges being divided into edges in different evidence sources in step F, a decision fusion method S2 is constructed to calculate the joint confidence m' e (x,y), the reliable weak edges are determined as edges according to the joint confidence, so as to obtain the final edge information of the image, and the edge connection is performed, specifically as follows:

[0141] Step G1: According to the probability of each pixel point being judged as an edge in steps F1, F2 and F3, a decision fusion method S2 is constructed to obtain the joint confidence m' e (x,y) of the weak edge pixel point.

[0142] m' e (x,y) = m 3e × m 4e × m 5e

[0143] Step G2: According to the joint confidence m' e (x,y) in step G1, it is determined whether the weak edge pixel point is an edge. For the weak edge pixel point (x,y), if the joint confidence m' e (x,y) corresponding to the weak edge pixel point is greater than a threshold pre, the weak edge pixel point is determined as an edge, otherwise, it is a non-edge, specifically as follows:

[0144]

[0145] Wherein, M2(x,y) is the edge of the image.

Claims

1. An evidence fusion driving-based water surface target edge detection algorithm, characterized in that, The method comprises the following steps: Step A: Collect the sample image of water surface target by visual sensor, and get the gray image I by graying the sample image gray (x, y); Step B: using Gaussian smoothing filter to suppress high-frequency noise in the gray image, so that the edge detection is more stable; Step C: constructing an edge feature extraction operator to extract the gradient amplitude M(x, y) and direction θ(x, y) of the image; Step D: constructing a "probabilistic non-maximum suppression mechanism" to make the edge determination credible, and recording the gradient amplitude M1(x, y) of the pixel point (x, y) that may be an edge; Step E: performing bilateral thresholding on the gradient amplitude M1(x, y) of the pixel point (x, y) to further divide the pixel point that may be an edge into a strong edge, a weak edge and a non-edge; Step F: determining the probability of dividing the weak edge into an edge from three aspects of gradient amplitude, gradient direction and neighborhood connectivity; Step G: According to the different evidence sources in step F, the weak edges are divided into the probability of edges, the decision fusion method S2 is constructed, and the joint confidence m' is calculated e (x, y), the reliable weak edges are determined as edges according to the joint confidence, and the final edge information of the image is obtained.

2. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The method for suppressing high-frequency noise in the gray image by using Gaussian smoothing filter in step B is as follows: Firstly, the Gaussian function S(x, y) is constructed, and the Gaussian smoothing filter is used to suppress the high-frequency noise in the gray image I gray (x, y), so that the edge detection is more stable, and the specific method is as follows: where σ is the standard deviation, which determines the width of the filter; then, taking the current pixel point as the coordinate origin, a Gaussian template S'(x, y) is constructed using the Gaussian function S(x, y); finally, the Gaussian template S'(x, y) is used to convolve the gray-scale image I gray (x, y) to obtain the smoothed image I1(x, y), as follows: I1(x,y) = I gray (x,y)*S'(x,y).

3. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The method for constructing an edge feature extraction operator to extract the gradient amplitude M(x, y) and direction θ(x, y) of the image in step C is as follows: Step Cl: Perform a difference operation on the smoothed image I1(x,y) to extract the x-direction gradient image G x (x,y) and y-direction gradient image G y (x,y), as follows: Step C2: Calculate the gradient magnitude M(x,y) and the gradient direction θ(x,y) of the image G x (x,y) and G y (x,y), the gradient magnitude M(x,y) and the gradient direction θ(x,y) of the image are calculated as follows:

4. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The method for constructing a "probabilistic non-maximum suppression mechanism" to make the edge determination credible in step D is as follows: Step D1: determining the class confidence m1 of the pixel point based on the similarity of the gradient direction and the maximum gradient in the neighborhood, specifically as follows: First, the gradient direction θ(x, y) of each point of the image calculated in step C2 is divided into an interval every 45 degrees, and the gradient direction of each point of the image after intervalization is recorded as θ'(x, y); Then, for any pixel point (x, y) in the image, a gradient amplitude matrix M(x, y) of adjacent pixels in the gradient direction θ'(x, y) corresponding to the pixel point is obtained, and the maximum gradient amplitude in M(x, y) is denoted as M max (x, y). Secondly, according to the similarity between M(x, y) and M max (x, y), the probability m 1e (x, y) that the point is an edge is calculated, as follows: d1(x,y) = M(x,y) - M max (x,y) where d1(x, y) represents the distance, i.e. the closeness, between M(x, y) and M max (x, y); m 1e (x, y) represents the probability that the pixel point is an edge pixel point; m 1n (x, y) represents the probability that the pixel point is a non-edge pixel point; and a represents a parameter adjustment factor. Step D2: constructing the class confidence m2 of the pixel point based on the neighborhood edge continuity, specifically as follows: First, for any pixel (x, y) in the image, count the m values ​​within its 3×3 neighborhood. 1e The number of pixels N1(x,y) whose (x±1,y±1) are greater than the confidence threshold thre1 is as follows: wherein f1(·) is an indicator function, when m 1e (x+i, y+j) > thre1, f1(·) is 1, otherwise 0; Then, the cascade probability m that the point is an edge is calculated from the number of pixels N1(x,y), as follows: 2e (x,y), as follows: where m 2e (x,y) represents the probability that the pixel point is an edge; m 2n (x,y) represents the probability that the pixel point is not an edge; Step D3: According to the probability of each pixel point being judged as an edge in steps D1 and D2, a credible edge decision fusion method S1 is constructed to obtain the fusion confidence m of each pixel point being judged as an edge e (x, y), specifically as follows: where m e (x,y) represents the fusion confidence that the pixel point is judged as an edge, m n (x,y) represents the fusion confidence that the pixel point is not an edge; Step D4: removing the pixel points that are not edges in the image according to the fusion confidence in step D3, combining the confidence threshold thre2, recording the gradient amplitude M1(x, y) of the pixel point that may be an edge, implementing "probabilistic non-maximum suppression", and refining the image edge, specifically as follows:

5. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The method for performing bilateral thresholding on the gradient amplitude M1(x, y) of step D to divide the edge into a strong edge, a weak edge and a non-edge in step E is as follows: wherein thre H wherein thre L denotes the gradient magnitude threshold; M2(x, y) = 255 indicates that the pixel point is a strong edge, M2(x, y) = 75 indicates that the pixel point is a weak edge, and M2(x, y) = 0 indicates that the pixel point is not an edge.

6. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The method for determining the probability of dividing the weak edge in step E into an edge from three evidence sources of gradient amplitude, gradient direction and neighborhood connectivity in step F is as follows: Step F1: Constructing the evidence source according to the gradient amplitude; from step E, the gradient amplitude of the weak edge is between the threshold thre H and thre L In addition, the greater the gradient amplitude of the current pixel point, the more likely it is to be an edge point. Under the design of the gradient amplitude evidence source, the probability of the weak edge pixel point being judged as an edge is as follows: where m 3e (x,y) denotes the probability that a weak edge pixel point (x,y) is judged as an edge; m 3n (x,y) denotes the probability that a weak edge pixel point (x,y) is judged as a non-edge. Step F2: constructing an evidence source according to the gradient direction; if the similarity of the gradient direction of the weak edge pixel point and the gradient direction of the strong edge pixel point in the neighborhood thereof is high, the probability that the weak edge pixel point is an edge is higher, and the specific steps are as follows: First, the number count4(x, y) of the strong edge pixel points in the neighborhood of the weak edge pixel point is counted, specifically as follows: where f2(·) is an indicator function that is 1 when M2(x+x i ,y+y i ) = 255 and 0 otherwise. Then, the similarity of the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in the neighborhood thereof is calculated, specifically as follows: Wherein, the higher the similarity of the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in the neighborhood thereof, the smaller sim4(x, y) is; Finally, according to count4(x, y) and sim4(x, y), the probability that the weak edge pixel point is judged as an edge from the gradient direction evidence source is as follows: where m 4e (x, y) represents the probability of the weak edge pixel point (x, y) being judged as an edge under the gradient direction evidence source; m 4n (x, y) represents the probability of the weak edge pixel point (x, y) being judged as a non-edge under the gradient direction evidence source; represents the evidence strength, that is, when there is a strong edge in the neighborhood of the weak edge pixel point, the greater the number of strong edges, the greater the probability of the weak edge pixel point being judged as a strong edge; represents the gradient direction similarity support term, that is, the higher the similarity of the gradient direction of the weak edge pixel point and the gradient direction of each strong edge pixel point in the neighborhood of the weak edge pixel point, the smaller sim4(x, y) is, the greater; when there is no strong edge in the neighborhood of the weak edge pixel point, the probability of the weak edge pixel point being judged as a strong edge is 0; Step F3: constructing evidence source according to neighborhood connectivity; according to edge continuity, if the gradient amplitude of the weak edge pixel point is similar to the gradient amplitude of the strong edge pixel point in its neighborhood, the probability that the weak edge pixel point is an edge is higher, and the specific steps are as follows: First, calculate the similarity of the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood, as follows: Wherein, the higher the similarity of the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood, the smaller sim5(x,y) is; Then, calculate the sum Nor(x,y) of the gradient amplitude calculated in step D4 corresponding to each strong edge pixel point in the neighborhood, as follows: Finally, according to count4(x,y), Nor(x,y) and sim5(x,y), determine the probability that the weak edge pixel point is judged as an edge under the neighborhood connectivity evidence source as follows: where m 5e (x, y) represents the probability that the weak edge pixel point (x, y) is judged as an edge under the neighborhood connectivity evidence source; m 5n (x, y) represents the probability that the weak edge pixel point (x, y) is judged as a non-edge under the neighborhood connectivity evidence source; represents the evidence strength, that is, when there is a strong edge in the neighborhood of the weak edge pixel point, the greater the number of strong edges, the greater the probability that the weak edge pixel point is judged as a strong edge; represents the gradient amplitude similarity support term, that is, the higher the similarity between the gradient amplitude of the weak edge pixel point and the gradient amplitude of each strong edge pixel point in its neighborhood, the smaller sim5(x, y) is, the greater.

7. The water surface target edge detection algorithm based on evidence fusion driving according to claim 1, characterized in that, The step G constructs a decision fusion method S2 according to the probability of dividing the weak edge into the edge under different evidence sources in the step F, and calculates the joint confidence m' e (x, y), the reliable weak edge is determined as the edge according to the joint confidence, so that the final edge information of the image is obtained, and the specific process is as follows: Step G1: According to the probability of each pixel point being judged as an edge in steps F1, F2 and F3, a decision fusion method S2 is constructed to obtain the joint confidence m' of the weak edge pixel point. e (x, y) is as follows: m' e (x,y) = m 3e × m 4e × m 5e Step G2: According to the joint confidence m' e (x,y), determine whether the weak edge pixel point is an edge; for the weak edge pixel point (x,y), if the joint confidence m' e (x,y) corresponding to the weak edge pixel point is greater than a threshold value pre, then the weak edge pixel point is determined to be an edge, otherwise it is a non-edge, as follows: Wherein, M2(x,y) is the edge of the image.

Images