Intelligent ship recognition and classification method based on multi-view collaborative SAR images

Abstract

The multi-view collaborative SAR image ship intelligent identification and classification method of the present application identifies multi-view SAR ships through information fusion evidence reasoning. First, the feature area of ​​interest of the ship is extracted based on the image interception and digging of bicubic interpolation. After dimensionality reduction through multi-view ship analysis, it is input into the ship identification classifier to obtain soft judgment and confidence probability output to form a corresponding evaluation matrix. Then, the improved evidence reasoning-assisted judgment is used to fuse the classification judgments under different viewpoints to obtain the final ship identification judgment result. It overcomes the limitations of previous single-view recognition, has strong system stability, reliability and anti-interference ability, greatly improves the accuracy of ship recognition, greatly improves the efficiency of ship recognition and classification in vast waters, helps the management services of waterways, ports, waters, etc., provides supporting basis and predictive decision-making, and has huge application value and economic benefits.

Description

Technical Field

[0001] The present application relates to a SAR image ship recognition and classification method, and in particular to a multi-view collaborative SAR image ship intelligent recognition and classification method, belonging to the technical field of SAR ship recognition and classification. Background Art

[0002] In recent years, multi-view pattern recognition technology has developed rapidly in the fields of machine learning and computer vision, and its application fields are becoming more and more extensive, ranging from face recognition, vehicle recognition to specific target recognition. The biggest difficulty in the recognition problem is how to identify the type of objects under multi-view observation, because different instances of the same object present different feature information in different time periods and environments, and the feature information presented by the same object is also different under different viewing angles. Therefore, multi-view target recognition must solve the two problems of intra-class differences and multi-view differences of the same target. At present, there are many mature methods for mathematical modeling of feature information of the same target, but there is still a lack of effective methods for mathematical models of multi-view of the same target. The main reason is that it is currently difficult to extract effectively associated feature information from multiple perspectives of the same target, which results in a decrease in recognition accuracy. Therefore, considering both intra-class features and multi-view factors at the same time, it is very challenging and valuable to realize a multi-view target recognition algorithm with a high recognition rate.

[0003] The development of synthetic aperture radar (SAR) pattern recognition technology has a history of several decades and has great practical significance in the field of image recognition and other related fields. SAR ship recognition and classification methods can be used for highly reliable recognition and classification of ship targets, accurate design and planning of waterway routes, etc. Due to the complex and diverse water environment, SAR images are urgently needed to have the ability to accurately identify ships in various environments.

[0004] In the management and service of waterways, ports, waters, etc., the rapid identification and accurate classification of ships is extremely important, especially the rapid and accurate classification of the basic characteristics and types of ships (such as cruise ships, cargo ships and fishing boats), which can provide supporting basis and predictive decision-making for the ship identification management of waterways and ports, and provide forward-looking support for their subsequent targeted management and services. At the same time, the accurate identification and classification of ships in the waterway can also provide technical support for the intelligent management of the waterway and the scientific planning of the route, and provide reliable guarantee for the comprehensive development and utilization of the waterway. SAR ship identification technology is one of the important supporting technologies for the management of waterways and ports today. Due to the increasingly changing water environment and the interference of various ships and complex environmental characteristics, the accurate identification of ships in complex waters has caused great obstacles.

[0005] As we all know, finding, identifying and classifying ships in vast waters (especially sea areas) is as difficult as looking for a needle in a haystack. Therefore, how to efficiently identify ships has become the key to current waterway navigation and port management. Intelligent identification and classification of ships in complex and vast waters is one of the major problems that need to be solved urgently today and for a long time to come. Therefore, it is of far-reaching practical significance to focus on observing ships at sea and develop SAR image ship fusion identification related technologies.

[0006] The existing SAR image ship recognition generally uses the relevant image processing algorithm to process the target area in the water area (sea area) to identify the type of ship it belongs to. Usually, the process mainly includes the following three steps: the first is preprocessing to extract the feature area of ​​interest, the second is feature extraction and transformation, and the third is classification and recognition.

[0007] The main functions of preprocessing are denoising and region segmentation, where denoising removes the interference of coherent speckle noise, and the role of region segmentation is to extract the feature region of interest. Usually, SAR image segmentation can be summarized into the following two types based on whether it can be segmented based on prior knowledge: data-driven method and model-driven method.

[0008] Feature extraction is to find out the relevant attribute information that can accurately express the category of the object after extracting the region of interest, reduce redundant information as much as possible, and improve the algorithm's running speed while ensuring accurate recognition. This is also the core content of SAR ship recognition. In order to accurately describe the recognition features of the target, the existing technology has done a lot of work on feature extraction. It is proposed to use the active appearance model to classify and recognize SAR image ships. The model includes two parts: shape and texture. The feature description of the target to be recognized is made from the two aspects of shape contour and grayscale features. The edge information is regarded as the ship recognition feature, and the target is classified and recognized through three-dimensional modeling and related modeling. Through the Hausdorff distance test of the peak signal, the second-order Gaussian function is used to fit the peak, and the position, width and amplitude of the peak signal are extracted.

[0009] Because SAR image data is a high-dimensional information vector, the amount of information processed by the recognition algorithm is often very large, which greatly reduces the recognition efficiency. Therefore, it is necessary to map the feature space of the image to a low-dimensional transformation space, namely, feature transformation. Feature transformation does not need to take into account any prior knowledge of the target features, has no obvious physical meaning, and only performs data processing and analysis in the field of linear algebra. Commonly used linear transformations include: linear discriminant analysis, wavelet transform, etc.

[0010] Information fusion in target recognition of existing technologies: After the classifier identifies the ships from different perspectives, it is necessary to achieve fusion recognition through information fusion technology. Information fusion integrates and associates data from different information sources at different levels to obtain a consistent description and definition of the observed target, so that the features of each part of the target can be more fully adopted to achieve more ideal decision-making and estimation. Generally, according to the level of image information processing in the information fusion process, it is divided into three categories: fusion at the data level, fusion at the feature level, and fusion at the decision level.

[0011] The first step of data layer fusion recognition is usually to associate and synthesize the input signals from non-single sensors, and the second step is to fuse the data, extract the key attribute data from the fused data results, and finally input them into the classifier to complete the classification recognition. The existing technology uses image information algebra to perform image information fusion methods from non-single perspectives. The data layer fusion level is the lowest, and its disadvantages are: first, the amount of information calculation is large, and there are certain requirements for bandwidth and computing power; second, because the original data information is used, it is easily affected by environmental factors; third, the matching accuracy of the input data participating in the fusion operation is required to be reduced to the level of one pixel, which invisibly increases the cost of operation.

[0012] Feature layer fusion recognition first extracts features from inputs from different perspectives, then fuses the feature vectors, and finally classifies the fusion results. Feature layer fusion recognition combines the advantages and disadvantages of the data layer and decision layer target recognition methods. The problem of different dimensions of feature vectors is more prominent during feature layer fusion recognition, and the dimension is very large, which creates obstacles for subsequent pattern recognition.

[0013] Decision-layer fusion recognition starts with preprocessing the input signals of each perspective, extracting feature data, and then making recognition decisions to obtain a preliminary unintegrated decision result. Then, the preliminary decision conclusions are fused through the fusion module to obtain the final target classification result. Decision-layer fusion does not require high transmission bandwidth and can well reflect the information of the observed objects in different scenes. In addition, it has low dependence on the source of information, becoming a hot spot in the current fusion recognition field.

[0014] In general, the existing SAR image ship recognition and classification technology has obvious deficiencies. Its main defects and design difficulties include:

[0015] First, the biggest difficulty in the existing target recognition problem is how to identify the type of objects under multi-view observation, because different instances of the same object present different feature information in different time periods and environments, and the feature information presented by the same object is also different under different viewing angles, so multi-view target recognition must solve the two problems of intra-class differences and multi-view differences of the same target. At present, there is still a lack of effective methods for the multi-view mathematical model of the same target, and it is difficult to extract effectively associated feature information under the multi-view of the same target, resulting in a decrease in the accuracy of recognition. Therefore, considering both intra-class features and multi-view factors at the same time, it is challenging and valuable to achieve multi-view ship recognition with a high recognition rate.

[0016] Second, in the management and services of waterways, ports, waters, etc., rapid identification and accurate classification of ships are extremely important. They can provide supporting basis and predictive decision-making for ship identification management in waterways and ports, and provide forward-looking support for their subsequent targeted management and services. They can also provide technical support for the intelligent management of waterways and the scientific planning of routes. However, due to the increasingly changing water environment and the interference of various ships and complex environmental characteristics, the accurate identification of ships in complex waters has caused great obstacles. Finding ships, identifying ships, and classifying ships in vast waters (especially sea areas) is tantamount to finding a needle in a haystack, which is very difficult. Therefore, how to efficiently identify ships has become the key to current waterway navigation and port management. Existing technologies cannot solve the problem of intelligent identification and classification of ships in complex and vast waters.

[0017] Third, the existing ship identification and classification technology generally uses effective feature data from a single perspective. Single-perspective identification has great limitations, and the system stability, reliability and anti-interference ability are relatively poor, resulting in low classification accuracy of ships at sea, and unable to effectively solve the SAR ship identification problem. The existing technology lacks an effective method to extract the feature area of ​​interest of the ship, and cannot be input into the ship identification classifier to obtain soft judgment and confidence probability output after dimensionality reduction through ship analysis. It is impossible to use the improved evidence reasoning to assist in the judgment to fuse the classification judgment under different perspectives to obtain the final ship identification judgment result. The accuracy of ship identification and classification is not enough, and the degree of intelligence is low, which directly leads to low efficiency, precision and accuracy of water (sea) management services;

[0018] Fourth, the existing technology lacks a method for extracting feature areas of interest of multi-perspective ships, a method for reducing the dimensionality of multi-perspective ships, a method for designing ship identification classifiers and outputting confidence probability, and a method for multi-perspective ship fusion identification based on evidence reasoning and decision-making. When there is a conflict of evidence, the traditional evidence reasoning and judgment method goes against common sense and integrates inaccurate fusion results. The confidence of the recognition results is low and shows a polarized trend. When the input image has speckle noise, the recognition effect is very poor, the recognition rate drops rapidly, and the anti-noise performance is poor. The existing technology cannot integrate the posterior probability information, and the accuracy of ship identification and classification is very poor. Summary of the invention

[0019] In view of the shortcomings of the prior art, this application identifies multi-view SAR ships through information fusion evidence reasoning, uses effective feature data obtained from different sensors or different perspectives, first extracts the feature area of ​​interest of the ship based on bicubic interpolation image interception and digging, and then performs dimensionality reduction through multi-view ship analysis and inputs it into the ship identification classifier to obtain soft judgment and confidence probability output to form a corresponding evaluation matrix, and then uses the improved evidence reasoning to assist in the judgment to fuse the classification judgment under different perspectives to obtain the final ship identification judgment result, which overcomes the limitations of previous single-view recognition, has strong system stability, reliability and anti-interference ability, greatly improves the accuracy of ship identification, is an important way to effectively solve the problem of SAR image ship identification, greatly improves the efficiency of ship identification and classification in vast waters, and greatly helps the management services of waterways, ports, waters, etc. It can provide supporting basis and predictive decision-making for ship identification management of waterways and ports, and provide forward-looking support for its subsequent targeted management and services, and has huge application value and economic benefits.

[0020] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0021] The intelligent ship recognition and classification method of multi-view collaborative SAR images first extracts the feature area of ​​interest of the ship based on the image interception and digging of bicubic interpolation. On this basis, the multi-view ship analysis is used to reduce the dimension and then input into the ship recognition classifier to obtain the soft judgment and confidence probability output to form the corresponding evaluation matrix. Then, the improved evidence reasoning auxiliary judgment is used to fuse the classification judgment under different viewpoints to obtain the final ship recognition judgment result.

[0022] The first step is to extract the feature regions of interest of ships from multiple perspectives. The feature regions of interest that may exist in the ship are determined from the source image obtained by the detector, and the interference of other image parts except the ship is eliminated. The feature regions of interest of the ship are extracted by image segmentation and block extraction. The process consists of four parts:

[0023] Process 1: Find the coordinates of each pixel position in the ship target image corresponding to the sampling point on the source image;

[0024] Process 2: Find the grayscale values ​​corresponding to the 16 adjacent positions of the sampling position D in the source image;

[0025] Process 3: The grayscale value of the sampling position D is calculated by the sum of the grayscale values ​​of 16 adjacent positions according to the weighted convolution;

[0026] Process 4: Assign the gray value of the sampling position D to the gray value of the corresponding point in the target image;

[0027] The second step is the dimensionality reduction method of multi-view ships: taking the gray value of the image pixel as the eigenvalue, an image with n rows and m columns has n×m dimensional eigenvalues, extracting the main influencing factors in the original feature vector, removing unnecessary redundant information, and reducing the dimensionality of the original image:

[0028] The third step is to design the ship identification classifier and output the confidence probability: the ship classification and identification adopts a three-classification model, which uses two two-classification modules to form a three-classification module, and then post-processes the SVM output results and converts them into posterior probabilities. Finally, the obtained soft decisions are used for fusion identification;

[0029] The fourth step is multi-perspective ship fusion identification based on evidence reasoning decision-making: introduce the expert confidence factor method, use evidence reasoning to recombine the new trust function, and improve the evidence reasoning-assisted judgment to obtain the ship fusion identification result.

[0030] Multi-view collaborative SAR image ship intelligent identification and classification method, further, multi-view ship interest feature area extraction: each ship has six sight angles, and each sight angle is divided into 19 rotation angles, that is, from 0 degrees to 180 degrees, each time the observation is made at an interval of 10 degrees. The most critical step is to find the source image position corresponding to the pixel point in the target area, which is achieved by using the image interception ship digging method;

[0031] First, θ in the target image is the known rotation angle of the multi-view observation. ChopWid and ChopHei are the width and height of the image after the cut block is set, which must include the entire ship. chopWid is a multiple of 4, which is the value of widthstep, indicating the bytes required to store a row of pixels.

[0032] Then, the coordinates of the ship's centroid in the source data are estimated, and their horizontal and vertical coordinates are marked as CentX and CentY respectively; the horizontal and vertical coordinates (i, j) of each point in the target image are recorded as dcol and drow, and the horizontal and vertical coordinates TempC and TempR of the sampling point D in the source image corresponding to each point in the target image are obtained by matrix operation:

[0033] TempC=cos(θ)×(dcol-0.5×ChopWid)+sin(θ)×(drow-0.5×ChopHei)+CentX Equation 1

[0034] TempR=-sin(θ)×(dcol-0.5*ChopWid)+cos(θ)×(drow-0.5×ChopHei)+CentY Equation 2

[0035] After obtaining the coordinates of the sampling position D, the bicubic interpolation method is used to obtain the grayscale value at D. The bicubic interpolation calculation includes 16 pixel points, where (i', j') represents the pixel coordinates of the pixel to be calculated in the source image, including the decimal part, dx represents the decimal coordinate in the X direction, and dy represents the decimal coordinate in the Y direction:

[0036]

[0037] Bicubic interpolation is the sum of the weighted convolution of the 16 pixels of the image as the new pixel value, where R(x) represents the interpolation expression and the distribution sampling expression is:

[0038]

[0039] The distributed sampling method is implemented based on cubic convolution calculation, which assigns D pixel values ​​to the corresponding pixel values ​​in the block image, and finally obtains the desired feature area of ​​interest.

[0040] Multi-view collaborative SAR image ship intelligent recognition and classification method, further, multi-view ship dimensionality reduction method: for a random vector X∈R n , whose mean value vector m x Covariance matrix C x for:

[0041] m x =E{X} Formula 5

[0042] C x =E{(xm x )(Xm x ) T Formula 6

[0043] X is an n-dimensional vector, (Xm x )(Xm x ) T Then it is an n×n matrix, C x The element C in ii is the variance of the i-th component of the X vector in the population vector, C x Element C ij Then the vector x i With x jThe covariance of the matrix C x is a real symmetric matrix; when x i With x j When they are independent, their covariance is equal to 0, and there is C ij =C ji =0, if k vectors are sampled from a random population, the mean vector is:

[0044]

[0045] The covariance matrix is:

[0046]

[0047] In the transformation, an orthogonal transformation A = R is defined n×n , let X∈R n The vector maps to Y∈R n denoted by , while making its components independent of each other:

[0048] Y=A(Xm x ) Formula 9

[0049] The covariance matrix of a random vector Y is a diagonal matrix:

[0050]

[0051] Among them C Y =E{(Ym Y )(Ym Y )},m Y =E{Y} is the mean vector of vector Y, C Y =AC x A T C x Diagonalize it and take all its rows as C x Then, all the eigenvectors are arranged from large to small according to the corresponding eigenvalues. Then, the first row in A is associated with the largest eigenvalue, and the last row in A is associated with the smallest eigenvalue. x is a real symmetric matrix, so there are n standard orthogonal eigenvectors forming A.

[0052] The components in Y are independent of each other. x Diagonalization is to decorrelate X, C x and C Y The eigenvalues ​​are the same, both are C Y Lambda on the diagonal j , and their eigenvectors are also consistent;

[0053] X can also be recovered from Y, according to the property of orthogonal vectors A -1 =AT , so find:

[0054] X=YA T +m X Formula 11

[0055] If C X The eigenvectors associated with the largest K eigenvalues ​​in the matrix construct a K×n matrix A, which reduces Y to K dimensions, and we get:

[0056]

[0057] X and The mean square error is:

[0058]

[0059] When K = n, the error is equal to 0, λ j Decreasingly, use the first K most important eigenvectors to reduce the error.

[0060] The intelligent ship recognition and classification method of multi-view collaborative SAR images, further, the specific process of the multi-view ship dimensionality reduction method is as follows:

[0061] Process 1: Rearrange the m×n training interest feature region images into m×n dimensional column vectors, calculate their mean vectors, and then use the mean vectors to center all samples;

[0062] Process 2: Calculate the diagonalized covariance matrix, calculate its eigenvalues, and then sort the associated eigenvectors from large to small according to the eigenvalues;

[0063] Process 3: Take the first K eigenvectors K≤n-1 to form a projection matrix A, project the training image obtained after centralization onto A, and obtain the image after dimensionality reduction;

[0064] Process 4: Operate the test data according to the above process to obtain the data after dimensionality reduction;

[0065] Process 5: Select a suitable classifier to classify and identify the test image.

[0066] Multi-view collaborative SAR image ship intelligent identification and classification method, further, ship identification classifier design and confidence probability output: after the dimension reduction of the target image in the feature area of ​​interest, the SVM classifier is used to classify the target, and then the SVM output result is post-processed and converted into a posteriori probability. Finally, the obtained soft judgment is used for fusion identification to obtain the ship identification result based on multi-view. The linear binary classification model with maximum interval in the feature space is used to find a classification hyperplane so that the distance between the data vectors belonging to two different classes and the plane is the largest, and the vertical length of the data points on both sides of the plane from the plane is the largest. The hyperplane with the maximum classification distance characteristic is calculated, and the concepts of function interval and geometric interval are introduced;

[0067] When the hyperplane wx+b=0 is known, the absolute value on the left side of the equal sign is used as the size of the interval between point x and the hyperplane. The correctness of the classification is determined by judging whether the positive and negative properties of wx+b are consistent with the positive and negative properties of the category label y. The relationship between y×(wx+b) and zero is used to judge the accuracy of the classification. The function interval is defined as:

[0068]

[0069] Assume that there is a point x in the feature space, and the point is projected vertically onto the hyperplane to obtain the corresponding point x0, and w is the normal vector of the plane. The distance from the sample x to the hyperplane is γ, then:

[0070]

[0071] At the same time, x0 is located in the hyperplane, so it makes f(x0) equal to 0. Substituting it into the equation w of the hyperplane T x+b=0, we get: w T x0+b=0, then let the formula Multiply both sides of the equal sign by w T , and it is also known that: w T x0 = -b and w T w=||W|| 2 ,get:

[0072]

[0073] To obtain the modulus of γ, multiply γ by its corresponding categorical attribute y, giving the definition of the geometric interval:

[0074]

[0075] The data points to be identified are classified. The farther the data points are from the hyperplane, the more reliable the confidence of the judgment result. To ensure the correctness of the judgment, the selected classification hyperplane makes the distance reach the farthest, which is half of the Gap value.

[0076] The objective function corresponding to the maximum margin classifier is defined as: maxγ, and the prerequisite is met:

[0077]

[0078] Convert the objective function into:

[0079]

[0080] The 1 / ||w|| in the above formula is the geometric interval, which satisfies: y(w T x+b)=1,y(w T x+b)>1.

[0081] The multi-view collaborative SAR image ship intelligent identification and classification method, further, adopts an indirect method to solve the three-classification problem, that is, two two-classification modules are used to form a three-classification module, and a soft judgment is proposed based on fuzzy sets. While giving the category judgment of the sample, the corresponding confidence of the sample for each category is calculated based on this, that is, its posterior probability.

[0082] Let f(x) be the classification result of SVM without threshold, that is, hard decision, and estimate SVM soft decision from hard decision, that is, posterior probability:

[0083]

[0084] In the above formula, A and B are the parameters to determine. Under the premise of ensuring the sparsity of SVM, the posterior probability is estimated and y i represents the category to which the sample point i belongs, and its value is +1 or -1, then y i Convert to t i :

[0085]

[0086] The values ​​of parameters A and B are estimated by the minimum negative log-likelihood of the sample points:

[0087] min-∑tilog(p i )+(1-t i )log(1-p i ) Formula 23

[0088] And p i satisfy:

[0089]

[0090] After obtaining the confidence probability output from the classifier, decision fusion is used to perform multi-perspective fusion recognition processing.

[0091] Multi-view collaborative SAR image ship intelligent identification and classification method, further, multi-view ship fusion identification based on evidence reasoning decision: the obtained SVM soft judgment results are combined with evidence reasoning assistance to realize the fusion identification algorithm under multi-view observation. If there is L in the whole process j The input sources of the sensors are preprocessed first to obtain the training results of each SVM, that is, the SVM posterior probability P j ,Finally, the evidence reasoning and evidence fusion are completed to obtain the final judgment of fusion identification.

[0092] Multi-view collaborative SAR image ship intelligent identification and classification method, further, ship fusion identification based on improved evidence reasoning assisted judgment: Assume that the confidence given by expert i is m i (A), the expert confidence factor is a i , giving the expert confidence factor a i The reliability after is m ai (A) and define:

[0093] m ai (A) = a i m i (A) Formula 25

[0094] a i =0 means that the expert i is completely untrustworthy. i =1 is the measure of a completely credible expert i. The expert confidence factor is in the interval (0,1) and has practical significance. i >a j , it means that compared with the information from expert j, the information from expert i is more credible, and the expert confidence factor reduces the impact of evidence conflicts in the actual judgment process by adding weights;

[0095] First set the membership function and normalize the data to construct the corresponding credibility function m j (A j ), when there is a conflict between the evidence feature sets, if one of the evidence is consistent with the majority of other evidence, it is considered that the fusion experiment result is greatly affected by this evidence; on the contrary, if one evidence is inconsistent with the majority of other evidence, the influence of this evidence on the final fusion result is minimized as much as possible, and the expert confidence factor corresponding to each piece of evidence is given;

[0096] First, define two evidence bodies m i The similarity of (A)(i=1,2) is Sim(m1(A),m2(A)):

[0097] Sim(m1(A), m2(A))=1-d(m1(A), m2(A)) Equation 26

[0098] The distance between two evidence bodies is represented by d(m1(A), m2(A)), and the following similarity matrix is ​​written for all evidence bodies:

[0099]

[0100] Define the support of the corresponding expert Sup(m i ,(A4)), i=1,…,K is the sum of each row in the similarity matrix, defining the credibility Crd of each expert i :

[0101]

[0102] Among them, we get:

[0103]

[0104] The expert confidence factor DCE i The calculation formula is as follows:

[0105]

[0106] After obtaining the expert confidence factor, the following is the expression of the confidence function given to the expert confidence factor:

[0107] m (DCE)i (A)=DCE i ×m i (A) Formula 31

[0108] Evidence reasoning is used to assist in the determination and recombination of new trust functions to obtain the final fusion result.

[0109] The intelligent ship recognition and classification method of multi-view collaborative SAR images, and further, the implementation process of ship fusion recognition based on improved evidence reasoning assisted judgment:

[0110] Step 1: Construct membership function and normalize data;

[0111] Step 2: Construct the evidence body similarity matrix and calculate the expert confidence factor;

[0112] Step 3: Multiply the expert confidence factor by the original evidence confidence;

[0113] Step 4: Achieve pairwise fusion of evidence bodies;

[0114] Step 5: The results of the pairwise fusion in step 4 are fused with the credibility of the next set of evidence, and the process is repeated continuously until the final fusion result is obtained.

[0115] Compared with the prior art, the innovations and advantages of this application are:

[0116] First, this application identifies multi-view SAR ships through information fusion evidence reasoning, using effective feature data obtained from different sensors or different perspectives, firstly extracting the feature area of ​​interest of the ship based on bicubic interpolation image interception and digging, and then inputting it into the ship identification classifier to obtain soft judgment and confidence probability output through multi-view ship analysis for dimensionality reduction, forming a corresponding evaluation matrix, and then using the improved evidence reasoning auxiliary judgment to fuse the classification judgment under different perspectives to obtain the final ship identification judgment result, overcoming the limitations of previous single-view recognition, having strong system stability, reliability and anti-interference ability, greatly improving the accuracy of ship identification, and being an important way to effectively solve the problem of SAR image ship identification, greatly improving the efficiency of ship identification and classification in vast waters, and greatly helping the management services of waterways, ports, waters, etc., and being able to provide supporting basis and predictive decision-making for ship identification management of waterways and ports, and provide forward-looking support for its subsequent targeted management and services, bringing huge economic benefits, and having great practical application value;

[0117] Second, in the first step of extracting the feature regions of interest of multi-perspective ships, a reliable image block segmentation method is designed to determine the feature regions of interest that may exist in the ship from the source image obtained by the detector, eliminate the interference of other image parts except the ship, and extract the feature regions of interest of the ship well through image segmentation and interception. The second step is the dimensionality reduction method of multi-perspective ships. Since the dimension of the existing feature region image information vector is too large, if it is directly input into the classifier for recognition, the amount of information that needs to be processed is very large, which greatly reduces the efficiency of the algorithm, and contains a large amount of redundant information. This redundant information is not helpful for improving the recognition rate. This application extracts the main influencing factors in the initial feature vector, removes unnecessary redundant data, and realizes the dimensionality reduction of the original image. On the premise of ensuring the recognition rate, it reduces the time taken for the program to run;

[0118] Third, the third step is the design of the ship identification classifier and the output of the confidence probability. The SVM classifier is used to classify and judge the target, and the difficulty of three-classification is overcome by the indirect method, that is, two two-classification modules are used to construct a three-classification module. The SVM output result is then converted into a posterior probability through the minimum negative log-likelihood estimation, and the recognition rate under different dimensions is tested. It is determined that when the dimension reduction dimension is 28 dimensions, as much redundant information as possible can be eliminated and the algorithm running time can be reduced. At the same time, a good recognition rate can be achieved, which is beneficial to support the fusion recognition process, and plays an important role in decision-making fusion of the posterior probability and improving the efficiency and accuracy of SAR image ship recognition classification.

[0119] Fourth, the fourth step is the multi-perspective ship fusion identification based on evidence reasoning decision. In order to solve the problem of fusion output errors when the evidence is greatly contradictory, the expert confidence factor method is introduced. It is mainly applicable to situations where there is conflicting evidence, and its judgment result is more accurate than the traditional evidence reasoning judgment. The original basic probability assignment is multiplied by the expert confidence factor to reduce the impact of the conflict, so as to reach a more appropriate judgment conclusion. The collective wisdom of multiple people is coordinated to solve the more complex ship identification and classification problems, and successfully solves the problem of increasingly complex water environment conditions and the explosion of SAR information data causing the processing volume to expand. The ship identification and classification decision time is greatly shortened, the anti-noise effect is good, and the artificial intelligence algorithm is fully utilized to build an intelligent decision-making model to build a stable, accurate, and high-speed intelligent ship identification and classification collective decision-making system, which has very practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 This is a schematic diagram of intelligent ship recognition and classification based on multi-view collaborative SAR images.

[0121] Figure 2 This is a schematic diagram of SAR ship images observed from different perspectives.

[0122] Figure 3 It is a schematic diagram of image segmentation, interception and block extraction of feature areas of interest of ships.

[0123] Figure 4 This is an example of image segmentation and block extraction of the feature area of ​​interest of a ship.

[0124] Figure 5 It is a schematic diagram of the geometric interval of the ship identification classifier design and confidence probability output.

[0125] Figure 6 It is a schematic diagram of the constraint relationship of the structured dynamic background adaptive linear programming model.

[0126] Figure 7 This is a statistical graph of the recognition accuracy of three types of ship test samples under three dimensionality reduction dimensions.

[0127] Figure 8 It is a schematic diagram of multi-view target fusion recognition assisted by SVM posterior probability and evidence reasoning.

[0128] Fig. 9 It is a flow chart for the implementation of ship fusion identification based on improved evidence reasoning assisted judgment.

[0129] Fig.10 It is a comparison and evaluation flowchart of the experimental results of evidence reasoning and judgment fusion before and after improvement.

[0130] Fig.11This is a schematic diagram of the anti-noise results of adding speckle noise to three types of SAR ship images. Specific implementation methods

[0131] The following, in conjunction with the accompanying drawings, further describes the technical solution of the multi-view collaborative SAR image ship intelligent identification and classification method proposed in the present application, so that those skilled in the art can better understand the present application and implement it.

[0132] For multi-view SAR ships, we first use the image interception and digging algorithm based on bicubic interpolation to extract the feature area of ​​interest of the ship, and then reduce the dimension and input it into the support vector machine classifier to obtain soft judgment, forming the corresponding evaluation matrix, and then use the improved evidence reasoning algorithm to fuse the classification judgments under different perspectives to obtain the final fusion recognition judgment. Through comprehensive analysis of experimental results, we can well extract the feature area of ​​interest, reduce the dimension in the selected dimension, and finally obtain a high multi-view fusion recognition rate, which well realizes the fusion recognition of ships based on SAR images.

[0133] The multi-view collaborative SAR image ship intelligent recognition and classification process in this application mainly consists of four parts: 1. preprocessing; 2. dimensionality reduction; 3. classification output; 4. recognition decision fusion based on evidence assistance. Figure 1 shown.

[0134] 1. Extraction of ship interest feature regions from multiple perspectives

[0135] The SAR ship images observed from different viewing angles are as follows Figure 2 Each ship has six sight angles (5 degree sight, 10 degree sight, 15 degree sight, 20 degree sight, 25 degree sight, 30 degree sight), and each sight angle is divided into 19 rotation angles, from 0 degree to 180 degree, and observation is performed at intervals of 10 degrees each time.

[0136] The purpose of preprocessing is to determine the possible interest feature areas of the ship from the source image, remove the interference of other image parts other than the ship as much as possible, and extract the interest feature areas of the ship through image segmentation and block extraction. The main process consists of four parts:

[0137] Process 1: Find the coordinates (including the decimal part) of each pixel position in the ship target image corresponding to the sampling point on the source image;

[0138] Process 2: Find the grayscale values ​​corresponding to the 16 adjacent positions of the sampling position D in the source image;

[0139] Process 3: The grayscale value of the sampling position D is calculated by the sum of the grayscale values ​​of 16 adjacent positions according to the weighted convolution;

[0140] Process 4: Assign the gray value of the sampling position D to the gray value of the midpoint of the corresponding target image (segmented image);

[0141] In the above four processes, the most critical step is to find the source image position corresponding to the pixel point in the target area, using Figure 3 The image shown captures the first step of the ship digging method.

[0142] First, θ in the target image is the known rotation angle of multi-view observation, ChopWid and ChopHei are the width and height of the image after the cut block is set, which must include the entire ship, where chopWid is a multiple of 4, which is the value of widthstep, indicating the bytes required to store a row of pixels; then, estimate the coordinates of the center of mass of the ship in the source data, and mark them as CentX and CentY respectively; record the horizontal and vertical coordinates (i, j) of each point in the target image as dcol and drow, and then the horizontal and vertical coordinates TempC and TempR of the sampling point D in the source image corresponding to each point in the target image are obtained by the matrix operation in the above formula:

[0143] TempC=cos(θ)×(dcol-0.5×ChopWid)+sin(θ)×(drow-0.5×ChopHei)+CentX Equation 1

[0144] TempR=-sin(θ)×(dcol-0.5*ChopWid)+cos(θ)×(drow-0.5×ChopHei)+CentY Equation 2

[0145] After obtaining the coordinates of the sampling position D, the bicubic interpolation method is used to obtain the grayscale value at D. The bicubic interpolation calculation includes 16 pixel points, where (i', j') represents the pixel coordinates of the pixel to be calculated in the source image, including the decimal part, dx represents the decimal coordinate in the X direction, and dy represents the decimal coordinate in the Y direction:

[0146]

[0147] Bicubic interpolation is the sum of the weighted convolution of the 16 pixels of the image as the new pixel value, where R(x) represents the interpolation expression and the distribution sampling expression is:

[0148]

[0149] The distributed sampling method is implemented based on the cubic convolution calculation. The D pixel value is assigned to the corresponding pixel value in the block image, and finally the desired feature area of ​​interest is obtained, such as Figure 4 shown.

[0150] 2. Dimensionality Reduction Method of Multi-View Ship

[0151] After extracting the feature region of interest from the image, the grayscale value of each point in the image is used as a feature attribute, then an image with n rows and m columns will have an n×m dimensional feature vector. According to the size of the existing feature region of interest image, the dimension of the information vector is too large, so if it is directly input into the classifier for recognition, the amount of information that needs to be processed is very large, which greatly reduces the efficiency of the algorithm, and it contains a large amount of redundant information, which is not helpful in improving the recognition rate.

[0152] Extract the main influencing factors in the original feature vector, remove unnecessary redundant information, simplify the dimension of the original image, and reduce the algorithm processing time while ensuring the recognition accuracy. For a random vector X∈R n , whose mean value vector m x Covariance matrix C x for:

[0153] m x =E{X} Formula 5

[0154] C x =E{(xm x )(Xm x ) T Formula 6

[0155] X is an n-dimensional vector, (Xm x )(Xm x ) T Then it is an n×n matrix, C x The element C in ii is the variance of the i-th component of the X vector in the population vector, C x Element C ij Then the vector x i With x j The covariance of the matrix C x is a real symmetric matrix; when x i With x j When they are independent, their covariance is equal to 0, and there is C ij =C ji =0, if k vectors are sampled from a random population, the mean vector is:

[0156]

[0157] The covariance matrix is:

[0158]

[0159] In the transformation, an orthogonal transformation A = R is defined n×n , let X∈R n The vector maps to Y∈R ndenoted by , while making its components independent of each other:

[0160] Y=A(Xm x ) Formula 9

[0161] The covariance matrix of a random vector Y is a diagonal matrix:

[0162]

[0163] Among them C Y =E{(Ym Y )(Ym Y )},m Y =E{Y} is the mean vector of vector Y, C Y =AC x A T C x Diagonalize it and take all its rows as C x Then, all the eigenvectors are arranged from large to small according to the corresponding eigenvalues. Then, the first row in A is associated with the largest eigenvalue, and the last row in A is associated with the smallest eigenvalue. x is a real symmetric matrix, so there are n standard orthogonal eigenvectors forming A.

[0164] The components in Y are independent of each other. x Diagonalization is to decorrelate X, C x and C Y The eigenvalues ​​are the same, both are C Y Lambda on the diagonal j , and their eigenvectors are also consistent;

[0165] X can also be recovered from Y, according to the property of orthogonal vectors A -1 =A T , so find:

[0166] X=YA T +m X Formula 11

[0167] If C X The eigenvectors associated with the largest K eigenvalues ​​in the matrix construct a K×n matrix A, which reduces Y to K dimensions, and we get:

[0168]

[0169] X and The mean square error is:

[0170]

[0171] When K = n, the error is equal to 0, λ jDescending, use the first K most important eigenvectors to reduce the error, the specific process is as follows:

[0172] Process 1: Rearrange the m×n training interest feature region images into m×n dimensional column vectors (or row vectors), calculate their mean vectors, and then use the mean vectors to center all samples;

[0173] Process 2: Calculate the diagonalized covariance matrix, calculate its eigenvalues, and then sort the associated eigenvectors from large to small according to the eigenvalues;

[0174] Process 3: Take the first K eigenvectors K≤n-1 to form a projection matrix A, project the training image obtained after centralization onto A, and obtain the image after dimensionality reduction;

[0175] Process 4: Operate the test data according to the steps described above to obtain the data after dimensionality reduction;

[0176] Process 5: Select a suitable classifier to classify and identify the test image.

[0177] 3. Ship identification classifier design and confidence probability output

[0178] After the dimension reduction of the target image in the feature area of ​​interest is performed, the SVM classifier is used to classify and judge the target, and then the SVM output result is post-processed and converted into a posterior probability. Finally, the obtained soft judgment is used for fusion recognition to obtain the ship recognition result based on multiple perspectives. The linear binary classification model with maximum interval in the feature space is used to find a classification hyperplane so that the distance between the data vectors belonging to two different classes and the plane is the largest, and the vertical length of the data points on both sides of the plane from the plane is the largest. The hyperplane with the maximum classification distance characteristic is calculated, and the concepts of function interval and geometric interval are introduced.

[0179] When the hyperplane wx+b=0 is known, the absolute value on the left side of the equal sign is used as the size of the interval between point x and the hyperplane. The correctness of the classification is determined by judging whether the positive and negative properties of wx+b are consistent with the positive and negative properties of the category label y. The relationship between y×(wx+b) and zero is used to judge the accuracy of the classification. The function interval is defined as:

[0180]

[0181] However, there is a contradiction in the function interval set above. When w and b are changed in the same proportion (for example, they are uniformly replaced by 2w and 2b), the value of the function interval f(x) is also changed to twice the original number (but the hyperplane is essentially the same as before). Therefore, the function interval alone cannot solve the problem. Suppose there is a point x in the feature space, and the point is projected vertically onto the hyperplane to obtain the corresponding point x0, and w is the normal vector of the plane. The distance from the sample x to the hyperplane is γ, as shown in Figure 5 As shown, we have:

[0182]

[0183] At the same time, x0 is located in the hyperplane, so it makes f(x0) equal to 0. Substituting it into the equation w of the hyperplane T x+b=0, we get: w T x0+b=0, then let the formula Multiply both sides of the equality sign by w T , and it is also known that: w T x0 = -b and w T w=||W|| 2 ,get:

[0184]

[0185] To obtain the modulus of γ, multiply γ by its corresponding categorical attribute y, giving the definition of the geometric interval:

[0186]

[0187] The data points to be identified are classified. The farther the data points are from the hyperplane, the more reliable the confidence of the judgment result. To ensure the correctness of the judgment, the selected classification hyperplane makes the distance reach the farthest, which is half of the Gap value.

[0188] The objective function corresponding to the maximum margin classifier is defined as: maxγ, and the prerequisite is met:

[0189]

[0190] Convert the objective function into:

[0191]

[0192] 1 / ||W|| in the above formula is the geometric interval, which satisfies: y(w T x+b)=1,y(w Tx+b)>1, this application considers that there are three types of ships, so it is a three-classification model, and adopts an indirect method to solve the three-classification problem, that is, two two-classification modules are used to form a three-classification module, and a soft judgment is proposed based on fuzzy sets. While giving the category judgment of the sample (hard judgment), the corresponding confidence of the sample for each category is calculated based on this, that is, its posterior probability.

[0193] Let f(x) be the classification result of SVM without threshold, that is, hard decision, and estimate SVM soft decision from hard decision, that is, posterior probability:

[0194]

[0195] In the above formula, A and B are the parameters to determine. Under the premise of ensuring the sparsity of SVM, the posterior probability is estimated and y i represents the category to which the sample point i belongs, and its value is +1 or -1, then y i Convert to t i :

[0196]

[0197] The values ​​of parameters A and B are estimated by the minimum negative log-likelihood of the sample points:

[0198] min-∑t i log(p i )+(1-t i )log(1-p i ) Formula 23

[0199] And p i satisfy:

[0200]

[0201] After obtaining the confidence probability output from the classifier, decision fusion is used to perform multi-perspective fusion recognition processing.

[0202] 1. Ship identification experimental results and evaluation

[0203] The data used in the experiment are SAR images of three types of ships, namely cruise ships, cargo ships and fishing boats. Each ship has six viewing angles, each viewing angle has 19 rotation angles, that is, each ship has 114 source image data. 60% of the source images are used as training data to train the SVM classifier, and the remaining 40% are used as test data.

[0204] When linear dimensionality reduction is performed, an average graph of each ship is obtained, which reflects the grayscale average of the samples involved in training or prediction at each point, such as Figure 6 shown.

[0205] A total of 171 training data of three types of ships are input into the SVM classifier for training processing. The training result is the model file. Then the prediction samples and the model file are input into the SVM classifier for prediction to obtain the posterior probability and classification results.

[0206] Before decision fusion, the recognition rates of samples in three groups (57 images in each group) with different dimensionality reduction dimensions were tested. The experimental statistical results of the correct recognition probabilities of the three types of ship test samples in three dimensionality reduction dimensions are shown in Figure 7 shown.

[0207] Depend on Figure 7 The experimental statistical results show that when the dimension reduction dimension is set to 28 dimensions, the recognition rates of cargo ships, cruise ships and fishing boats are relatively high. When the dimension is reduced to 50 dimensions, although the recognition rate of cargo ships is very high, the recognition rate of cruise ships is too low, and the recognition rate of fishing boats is not high. When the dimension is reduced to 70 dimensions, although the recognition rates of cruise ships and fishing boats are very high, the recognition rate of cargo ships is too low. Based on the above analysis, when the dimension reduction dimension is selected to be 28 dimensions, as much redundant information as possible can be eliminated, the algorithm running time can be reduced, and a good recognition rate can be achieved at the same time. Therefore, in the following fusion recognition process, this application selects the 28-dimensional dimensionality reduction result to perform decision fusion on the posterior probability.

[0208] 4. Multi-perspective ship fusion identification based on evidence-based reasoning decision-making

[0209] The obtained SVM soft judgment results are combined with evidence reasoning assistance to realize the fusion recognition algorithm under multi-view observation. If there is L in the whole process j The input sources of the sensors are preprocessed first to obtain the training results of each SVM, that is, the SVM posterior probability P j ,Finally, the evidence reasoning and evidence fusion are completed to obtain the final judgment of fusion recognition. The schematic diagram of SVM posterior probability and evidence reasoning assisted multi-view target fusion recognition is shown in Figure 8 shown.

[0210] 1. Purpose and significance of fusion judgment and recognition

[0211] Group decision-making uses the collective wisdom of multiple people to solve more complex real-world problems. Today, the decision-making environment and conditions are becoming more and more complex. As the daily learning production rate increases, the decision-making time is also required to be shorter and shorter. The explosion of information data has led to an increasing amount of information processing. Based on this changing condition, how to fully utilize ML and artificial intelligence frontier research and cutting-edge technologies to build intelligent decision-making models and construct a stable, accurate, and high-speed intelligent collective decision-making system has very practical application value.

[0212] Identifying multi-view SAR ships through information fusion can use effective feature information from different sensors or different viewpoints, overcoming the limitations of previous single-view recognition. It has strong system stability, reliability and anti-interference ability, greatly improving the accuracy of ship recognition, and is an important way to effectively solve the problem of ship recognition in multi-view SAR images.

[0213] (II) Ship fusion identification based on improved evidence reasoning assisted judgment

[0214] In the absence of prior knowledge, evidence reasoning fusion is a good way to deal with uncertain types of reasoning, and the evidence reasoning synthesis rules are commutative and associative. However, the defect of this method is that when the evidence is highly contradictory, the fusion result is very different from the actual situation. In order to solve the problem of fusion output errors when the evidence is highly contradictory, this application introduces the expert confidence factor method.

[0215] Assume that the confidence given by expert i is m i (A), the expert confidence factor is a i , giving the expert confidence factor a i The reliability after is m ai (A) and define:

[0216] m ai (A) = a i m i (A) Formula 25

[0217] a i =0 means that the expert i is completely untrustworthy. i =1 is the measure of a completely credible expert i. The expert confidence factor is in the interval (0,1) and has practical significance. i >a j , it means that compared with the information from expert j, the information from expert i is more credible, and the expert confidence factor reduces the impact of evidence conflicts in the actual judgment process by adding weights.

[0218] First set the membership function and normalize the data to construct the corresponding credibility function m j (A j ), when there is a conflict between evidence feature sets, if one of the evidence is consistent with most of the other evidence, it is considered that the fusion experiment result is greatly affected by this evidence; on the contrary, if one evidence is inconsistent with most of the other evidence, the influence of this evidence on the final fusion result is minimized as much as possible, and an expert confidence factor corresponding to each evidence is given to eliminate adverse effects.

[0219] First, define two evidence bodies m iThe similarity of (A)(i=1,2) is Sim(m1(A),m2(A)):

[0220] Sim(m1(A), m2(A))=1-d(m1(A), m2(A)) Equation 26

[0221] The distance between two evidence bodies is represented by d(m1(A), m2(A)), and the following similarity matrix is ​​written for all evidence bodies:

[0222]

[0223] Define the support of the corresponding expert Sup(m i ,(A4)), i=1,…,K is the sum of each row in the similarity matrix, defining the credibility Crd of each expert i :

[0224]

[0225] Among them, we get:

[0226]

[0227] The expert confidence factor DCE i The calculation formula is as follows:

[0228]

[0229] After obtaining the expert confidence factor, the following is the expression of the confidence function given to the expert confidence factor:

[0230] m (DCE)i (A)=DCE i ×m i (A) Formula 31

[0231] By using evidence reasoning to assist in judgment and recombining the new trust function, we can get the final fusion result.

[0232] In summary, the implementation flow chart of ship fusion identification based on improved evidence reasoning assisted judgment is as follows: Fig. 9 :

[0233] Step 1: Construct membership function and normalize data;

[0234] Step 2: Construct the evidence body similarity matrix and calculate the expert confidence factor;

[0235] Step 3: Multiply the expert confidence factor by the original evidence credibility;

[0236] Step 4: Achieve pairwise fusion of evidence bodies;

[0237] Step 5: The results of the pairwise fusion in step 4 are fused with the credibility of the next set of evidence, and the process is repeated continuously until the final fusion result is obtained.

[0238] (III) Comparison and evaluation of the experimental results of evidence reasoning and judgment fusion before and after improvement

[0239] The experimental data used in this paper are multi-view SAR ship images. Through the previous processing, the SVM classifiers corresponding to the two sets of data and the pre-processed test samples to be identified have been obtained. The use and distribution of all experimental data are shown in the following figure. Fig.10 shown.

[0240] For the test images of SAR image ships, four groups of experiments are conducted:

[0241] The first set of experiments: All 57 test samples were input into the evidence reasoning fusion device for fusion of posterior probabilities. The experimental results show that the posterior probability value after traditional evidence reasoning fusion has serious errors, and the value before fusion is completely invisible. The posterior probability of the improved evidence reasoning fusion is basically the same as the posterior probability of each expert before fusion, and the posterior probability of the correct ship has been significantly improved. The analysis and statistical results have the following conclusions:

[0242] 1. When there is conflicting evidence, the traditional evidence reasoning method will violate common sense and generally result in inaccurate fusion results.

[0243] 2. The improved evidence reasoning judgment is an improvement on the traditional evidence reasoning judgment. It is mainly applicable to situations where there is conflicting evidence. Its judgment result is more accurate than the traditional evidence reasoning judgment. It reduces the impact of the conflict by multiplying the original basic probability assignment by the expert confidence factor, thereby reaching a more appropriate judgment conclusion.

[0244] The second set of experiments: Select images from six angles of the cruise ship test data for soft judgment and improve the evidence reasoning fusion, and record the results. Through the fusion results of the six angles, it can be concluded that the confidence of the correct recognition result is greater, while the confidence of the wrong recognition result is smaller, showing a polarized trend.

[0245] The third group of experiments: 57 test images are grouped into groups of three, so there are 29,260 possible permutations and combinations. Randomly select 200 of them for improved evidential reasoning and judgment fusion. Count the recognition outputs of each group, and the type with the largest posterior probability is the recognition result after fusion. Through the recognition statistical results, it can be concluded that the recognition result obtained by fusing the posterior probability using the evidential reasoning decision theory has a greatly improved recognition rate compared with the simple SVM classification with a single perspective, because it integrates the posterior probability information from multiple angles, so it can achieve a good multi-perspective target recognition effect.

[0246] The fourth group of experiments: adding speckle noise to the SAR ship image after extracting the feature region of interest, testing the anti-noise performance of the fusion recognition algorithm of this application. The feature region of interest after adding Gaussian noise is as follows: Fig.11 shown.

[0247] The improved evidence reasoning fusion recognition results are statistically calculated for the feature area of ​​interest after adding speckle noise according to the above steps. It can be concluded from the recognition rate statistics that the multi-view SAR ship fusion recognition algorithm adopted in this application still has good recognition effect when the input image has speckle noise, and the recognition rate is still above 93%, which has good noise resistance performance.

Claims

1. A method for intelligent ship recognition and classification based on multi-view collaborative SAR images, characterized in that: Firstly, the image interception and digging of bicubic interpolation is used to extract the feature area of ​​interest of the ship. On this basis, the dimension is reduced through multi-view ship analysis and then input into the ship recognition classifier to obtain soft judgment and confidence probability output to form the corresponding evaluation matrix. Then, the improved evidence reasoning auxiliary judgment is used to fuse the classification judgment under different perspectives to obtain the final ship recognition judgment result. The first step is to extract the feature regions of interest of ships from multiple perspectives. The feature regions of interest that may exist in the ship are determined from the source image obtained by the detector, and the interference of other image parts except the ship is eliminated. The feature regions of interest of the ship are extracted by image segmentation and block extraction. The process consists of four parts: Process 1: Find the coordinates of each pixel position in the ship target image corresponding to the sampling point on the source image; Process 2: Find the grayscale values ​​corresponding to the 16 adjacent positions of the sampling position D in the source image; Process 3: The grayscale value of the sampling position D is calculated by the sum of the grayscale values ​​of 16 adjacent positions according to the weighted convolution; Process 4: Assign the gray value of the sampling position D to the gray value of the corresponding point in the target image; The second step is the dimensionality reduction method of multi-view ships: taking the gray value of the image pixel as the eigenvalue, an image with n rows and m columns has n×m dimensional eigenvalues, extracting the main influencing factors in the original feature vector, removing unnecessary redundant information, and reducing the dimensionality of the original image: The third step is to design the ship identification classifier and output the confidence probability: the ship classification and identification adopts a three-classification model, which uses two two-classification modules to form a three-classification module, and then post-processes the SVM output results and converts them into posterior probabilities. Finally, the obtained soft decisions are used for fusion identification; The fourth step is multi-perspective ship fusion identification based on evidence reasoning decision-making: introducing the expert confidence factor method, using evidence reasoning to re-combine the new trust function, and improving the evidence reasoning-assisted judgment to obtain the ship fusion identification result; Extraction of interest feature regions of multi-view ships: Each ship has six sight angles, and each sight angle is divided into 19 rotation angles, that is, from 0 to 180 degrees, and observation is performed at intervals of 10 degrees each time. The most critical step is to find the source image position corresponding to the pixel point in the target area, which is achieved by using the image interception ship digging method; First, θ in the target image is the known rotation angle of the multi-view observation. ChopWid and ChopHei are the width and height of the image after the cut block is set, which must include the entire ship. chopWid is a multiple of 4, which is the value of widthstep, indicating the bytes required to store a row of pixels. Then, the coordinates of the ship's centroid in the source data are estimated, and their horizontal and vertical coordinates are marked as CentX and CentY respectively; the horizontal and vertical coordinates (i, j) of each point in the target image are recorded as dcol and drow, and the horizontal and vertical coordinates TempC and TempR of the sampling point D in the source image corresponding to each point in the target image are obtained by matrix operation: TempC=cos(θ)×(dcol-0.5×ChopWid)+sin(θ)×(drow-0.5×ChopHei)+CentX Equation 1 TempR=-sin(θ)×(dcol-0.5*ChopWid)+cos(θ)×(drow-0.5×ChopHei)+CentY Equation 2 After obtaining the coordinates of the sampling position D, the bicubic interpolation method is used to obtain the grayscale value at D. The bicubic interpolation calculation includes 16 pixel points, where (i', j') represents the pixel coordinates of the pixel to be calculated in the source image, including the decimal part, dx represents the decimal coordinate in the X direction, and dy represents the decimal coordinate in the Y direction: Bicubic interpolation is the sum of the weighted convolution of the 16 pixels of the image as the new pixel value, where R(x) represents the interpolation expression and the distribution sampling expression is: The distributed sampling method is implemented based on cubic convolution calculation, which assigns D pixel values ​​to the corresponding pixel values ​​in the block image, and finally obtains the desired feature area of ​​interest; Ship identification classifier design and confidence probability output: After the dimension of the target image in the feature area of ​​interest is reduced, the SVM classifier is used to classify the target, and then the SVM output result is post-processed and converted into a posterior probability. Finally, the obtained soft judgment is used for fusion recognition to obtain the ship identification result based on multiple perspectives. The linear binary classification model with maximum interval in the feature space is used to find a classification hyperplane so that the distance between the data vectors belonging to two different classes and the plane is the largest, and the vertical length of the data points on both sides of the plane from the plane is the largest. The hyperplane with the maximum classification distance characteristic is calculated, and the concepts of function interval and geometric interval are introduced; When the hyperplane wx+b=0 is known, the absolute value on the left side of the equal sign is used as the size of the interval between point x and the hyperplane. The correctness of the classification is determined by judging whether the positive and negative properties of wx+b are consistent with the positive and negative properties of the category label y. The relationship between y×(wx+b) and zero is used to judge the accuracy of the classification. The function interval is defined as: Assume that there is a point x in the feature space, and the point is projected vertically onto the hyperplane to obtain the corresponding point x0, and w is the normal vector of the plane. The distance from the sample x to the hyperplane is γ, then: At the same time, x0 is located in the hyperplane, so it makes f(x0) equal to 0. Substituting it into the equation w of the hyperplane T x+b=0, we get: w T x0+b=0, then let the formula Multiply both sides of the equal sign by w T , and it is also known that: w T x0 = -b and w T w=||W|| 2 ,get: To obtain the modulus of γ, multiply γ by its corresponding categorical attribute y, giving the definition of the geometric interval: The data points to be identified are classified. The farther the data points are from the hyperplane, the more reliable the confidence of the judgment result. To ensure the correctness of the judgment, the selected classification hyperplane makes the distance reach the farthest, which is half of the Gap value. The objective function corresponding to the maximum margin classifier is defined as: maxγ, and the prerequisite is met: Convert the objective function into: 1 / ||W|| in the above formula is the geometric interval, which satisfies: y(w T x+b)=1,y(w T x+b)>1.

2. According to the multi-view collaborative SAR image ship intelligent identification and classification method of claim 1, it is characterized by: Dimensionality reduction method for multi-view ships: For a random vector X∈R n , whose mean value vector m x Covariance matrix C x for: m x =E{X} Formula 5 X is an n-dimensional vector, (Xm x )(Xm x ) T Then it is an n×n matrix, C x The element C in ii is the variance of the i-th component of the X vector in the population vector, C x Element C ij Then the vector x i With x j The covariance of the matrix C x is a real symmetric matrix; when x i With x j When they are independent, their covariance is equal to 0, and there is C ij =C ji =0, if k vectors are sampled from a random population, the mean vector is: The covariance matrix is: In the transformation, an orthogonal transformation A = R is defined n×n , let X∈R n The vector maps to Y∈R n denoted by , while making its components independent of each other: Y=A(Xm x ) Formula 9 The covariance matrix of a random vector Y is a diagonal matrix: Among them C Y =E{(Ym Y )(Ym Y )},m Y =E{Y} is the mean vector of vector Y, C Y =AC x A T C x Diagonalize it and take all its rows as C x Then, all the eigenvectors are arranged from large to small according to the corresponding eigenvalues. Then, the first row in A is associated with the largest eigenvalue, and the last row in A is associated with the smallest eigenvalue. x is a real symmetric matrix, so there are n standard orthogonal eigenvectors forming A; The components in Y are independent of each other. x Diagonalization is to decorrelate X, C x and C Y The eigenvalues ​​are the same, both are C Y Lambda on the diagonal j , and their eigenvectors are also consistent; X can also be recovered from Y, according to the property of orthogonal vectors A -1 =A T , so find: X = YA T +m X Equation 11 If C X The eigenvectors associated with the largest K eigenvalues ​​in the matrix construct a K×n matrix A, which reduces Y to K dimensions, and we get: X and The mean square error is: When K = n, the error is equal to 0, λ j Decreasingly, use the first K most important eigenvectors to reduce the error.

3. According to the multi-view collaborative SAR image ship intelligent identification and classification method of claim 2, it is characterized by: The specific process of the multi-view ship dimensionality reduction method is as follows: Process 1: Rearrange the m×n training interest feature region images into m×n dimensional column vectors, calculate their mean vectors, and then use the mean vectors to center all samples; Process 2: Calculate the diagonalized covariance matrix, calculate its eigenvalues, and then sort the associated eigenvectors from large to small according to the eigenvalues; Process 3: Take the first K eigenvectors K≤n-1 to form a projection matrix A, project the training image obtained after centralization onto A, and obtain the image after dimensionality reduction; Process 4: Operate the test data according to the above process to obtain the data after dimensionality reduction; Process 5: Select a suitable classifier to classify and identify the test image.

4. The method for intelligent ship recognition and classification based on multi-view collaborative SAR images according to claim 1 is characterized in that: An indirect method is used to solve the three-classification problem, that is, two two-classification modules are used to form a three-classification module. A soft decision is proposed based on fuzzy sets. While giving the category decision of the sample, the corresponding confidence of the sample for each category is calculated, that is, its posterior probability. Let f(x) be the classification result of SVM without threshold, that is, hard decision, and estimate SVM soft decision from hard decision, that is, posterior probability: In the above formula, A and B are the parameters to determine. Under the premise of ensuring the sparsity of SVM, the posterior probability is estimated and y i represents the category to which the sample point i belongs, and its value is +1 or -1, then y i Convert to t i : The values ​​of parameters A and B are estimated by the minimum negative log-likelihood of the sample points: min-Σt i log(p i )+(1-t i )log(1-p i ) Equation 23 And p i satisfy: After obtaining the confidence probability output from the classifier, decision fusion is used to perform multi-perspective fusion recognition processing.

5. The multi-view collaborative SAR image ship intelligent identification and classification method according to claim 1 is characterized by: Multi-view ship fusion recognition based on evidence reasoning decision: The obtained SVM soft judgment results are combined with evidence reasoning assistance to realize the fusion recognition algorithm under multi-view observation. If there is L in the whole process j The input sources of the sensors are preprocessed first to obtain the training results of each SVM, that is, the SVM posterior probability P j ,Finally, the evidence reasoning and evidence fusion are completed to obtain the final judgment of fusion identification.

6. The multi-view collaborative SAR image ship intelligent identification and classification method according to claim 1 is characterized by: Ship fusion identification based on improved evidence reasoning assisted judgment: Assume that the confidence given by expert i is m i (A), the expert confidence factor is a i , giving the expert confidence factor a i The reliability after is ma i (A) and define: m ai (A) = a i m i (A) Formula 25 a i =0 means that the expert i is completely untrustworthy. i =1 is the measure of a completely credible expert i. The expert confidence factor is in the interval (0,1) and has practical significance. i >a j , it means that compared with the information from expert j, the information from expert i is more credible, and the expert confidence factor reduces the impact of evidence conflicts in the actual judgment process by adding weights; First set the membership function and normalize the data to construct the corresponding credibility function m j (A j ), when there is a conflict between the evidence feature sets, if one of the evidence is consistent with the majority of other evidence, it is considered that the fusion experiment result is greatly affected by this evidence; on the contrary, if one evidence is inconsistent with the majority of other evidence, the influence of this evidence on the final fusion result is minimized as much as possible, and the expert confidence factor corresponding to each piece of evidence is given; First, define two evidence bodies m i Similarity Sim(m1(A), m2(A)) of (A)(i=1,2): Sim(m1(A), m2(A))=1-d(m1(A), m2(A)) Equation 26 The distance between two evidence bodies is represented by d(m1(A), m2(A)), and the following similarity matrix is ​​written for all evidence bodies: Define the support of the corresponding expert Sup(m i , (A4)), i = 1, ..., K is the sum of each row in the similarity matrix, and defines the credibility Crd of each expert i : Among them, we get: The expert confidence factor DCE i The calculation formula is as follows: After obtaining the expert confidence factor, the following is the expression of the confidence function given to the expert confidence factor: m (DCE)i (A) = DCE i × m i (A) Equation 31 Evidence reasoning is used to assist in the determination and recombination of new trust functions to obtain the final fusion result.

7. The method for intelligent ship recognition and classification based on multi-view collaborative SAR images according to claim 6 is characterized in that: The implementation process of ship fusion identification based on improved evidence reasoning assisted judgment: Step 1: Construct membership function and normalize data; Step 2: Construct the evidence body similarity matrix and calculate the expert confidence factor; Step 3: Multiply the expert confidence factor by the original evidence confidence; Step 4: Achieve pairwise fusion of evidence bodies; Step 5: The results of the pairwise fusion in step 4 are fused with the credibility of the next set of evidence, and the process is repeated continuously until the final fusion result is obtained.

Images