Ship classification method based on HRRP and SAR data trusted decision fusion

By fusing the posterior probability and entropy of deep learning model, the redistribution of initial deletion confidence function by fusion of evidence theory, the problem of low accuracy and poor interpretability of single data sources in ship classification is solved, and higher classification accuracy and ability Reliability assessment.

CN118587549BActive Publication Date: 2024-11-15HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411069633.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2024-08-05
Filing Date
2024-08-06
Publication Date
2024-11-15
Estimated Expiration
2044-08-06

AI Technical Summary

Technical Problem

When the prior art uses a single SAR image and HRRP data in ship classification, the classification accuracy is low and the robustness is poor, and the interpretability of the deep learning model is poor, making it impossible to estimate the credibility of the classification results.

Method used

The evidence theory method is used to fuse the sample posterior probability of deep learning models, introduce classification accuracy as prior knowledge, redistribute the initial confidence function with entropy, calculate the credibility of samples and classification results, and output the classification results. Reliability value.

Benefits of technology

It improves the accuracy and robustness of ship classification, provides an estimate of the credibility of classification results, and enhances the interpretability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118587549B_ABST
    Figure CN118587549B_ABST
Patent Text Reader

Abstract

The present invention discloses a ship classification method based on the fusion of HRRP and SAR data trustworthy decision. The method first obtains the ship HRRP and SAR image data and performs preprocessing respectively. Secondly, two classification models are constructed, and supervised training is performed on the SAR image data and HRRP data using training sets respectively, and the trust values ​​A1 and A2, as well as the corresponding posterior probabilities and classification prediction results are obtained respectively. Then, the two posterior probabilities obtained are redistributed, and two new sets of initial confidence functions are constructed. The two sets of initial confidence functions are fused to obtain the classification results of the target samples and the fused trust values ​​A3 and A4. Finally, A3, A4, A1, and A2 are AA-fused and calculated, and the trust corresponding to the classification result is output. The present invention improves the classification accuracy, the reliability and robustness of the initial confidence function, and the accuracy of ship classification when there are many target categories.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a ship classification method based on HRRP and SAR data trusted decision fusion. Background Art

[0002] Synthetic aperture radar (SAR) and one-dimensional range profile (HRRP) are both high-resolution radar data. SAR is an active ground observation system that can image targets in the presence of cover and low visibility. The image reflects the geometric characteristics and scattering characteristics of the target, and has the advantages of all-day, all-weather, and wide detection range. HRRP is obtained by high-resolution radar. When the target size is much larger than the radar resolution unit size, the radar echo of the target forms HRRP, which reflects the geometric structure characteristics of the target in the distance direction, including the size of the target, the position of the scattering center, etc. It has the advantages of small data volume, good real-time performance, easy processing, and strong anti-interference ability. Target classification methods based on HRRP have become a research hotspot in recent years.

[0003] SAR images usually have poor image quality and severe coherent speckle noise, and after filtering, the details of the ship are seriously lost. HRRP data contains more detailed information about ships, but also has the problem of azimuth sensitivity. Therefore, the accuracy of ship classification using a single SAR image and HRRP data is difficult to meet actual requirements and has poor robustness.

[0004] In the field of ship classification, with the development of deep learning, the advantages of deep learning-based classification methods are becoming increasingly significant. By using nonlinear network structures to effectively extract and learn the features of input data, the classification of ships is completed. It has good feature extraction and learning capabilities, and effectively makes up for the shortcomings of artificially designed feature extraction in traditional methods. However, one problem with deep learning models is poor interpretability. After outputting the classification results, people cannot make a reasonable estimate of the credibility of the classification results through the analysis process, that is, they cannot know whether the classification results are credible. Summary of the invention

[0005] The purpose of the present invention is to address the defects of the prior art and propose a ship classification method based on the fusion of HRRP and SAR data trustworthy decision. The method uses the evidence theory method to fuse the sample posterior probability output by the trained deep learning model, introduces the prior knowledge of the classification accuracy of the deep learning model, and redistributes the initial confidence function in combination with entropy to improve the classification accuracy after fusion, calculates the sample credibility at each stage, and finally outputs the credibility value of the classification result.

[0006] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0007] The ship classification method based on HRRP and SAR data trusted decision fusion includes the following steps:

[0008] S101: Acquire HRRP one-dimensional data and SAR image data of the same ship target at the same time and corresponding labels to obtain an original data set, pre-process the HRRP one-dimensional data and SAR image detection data respectively, and divide them into a training set and a test set.

[0009] S102: Construct a CNN classification model to perform supervised training on the SAR image data, use the training set for training, record the relevant credibility value A1, and predict the test set to give the posterior probability of each sample being classified into a certain category and the classification prediction result.

[0010] S103: Construct an MLP classification model to perform supervised training on the HRRP data, use the training set for training, record the relevant credibility value A2, and predict the test set to give the posterior probability of each sample being classified into a certain category and the classification prediction result.

[0011] S104: construct an initial confidence allocation module for redistributing the posterior probabilities obtained in S102 and S103, and construct two new sets of initial confidence functions derived from SAR data and HRRP data.

[0012] S105: Construct an evidence theory fusion module to fuse the two sets of initial belief functions to obtain the classification results of the target samples, and record the fused credibility values ​​A3 and A4.

[0013] S106: Input the posterior probabilities saved in S102 and S103 into the S104 initial confidence assignment module and the S105 evidence theory fusion module in sequence to obtain the fused test set classification result, calculate the credibility values ​​A3 and A4 of the fused sample classification, and perform AA fusion calculation with the previously recorded category credibility values ​​A1 and A2, and output the credibility corresponding to the classification result.

[0014] The data preprocessing in S101 is specifically as follows:

[0015] The SAR image is filtered by a Butterworth low-pass filter to enhance the contrast of the target area, making the features more obvious and convenient for training. The preprocessing steps can be expressed as follows:

[0016]

[0017]

[0018] in, is an image The pixel coordinates in is the filter kernel The element coordinates in is the distance between the midpoint of the frequency domain and the center of the frequency rectangle. The cutoff frequency is located at a distance from the origin. Where The transfer function of the 2nd-order Buttworth low-pass filter is , This is the image after filtering.

[0019] For HRRP data, the sampling feature extraction method is used for preprocessing to keep the original sequence characteristics and overall change rules as much as possible. The specific implementation is as follows:

[0020]

[0021]

[0022]

[0023] in, represents the original HRRP data, Indicates the total number of distance units contained in the HRRP data. is the HRRP data output by preprocessing. The total number of distance units after processing is , is the sampling extraction method. The total number of categories that need to be classified is , represent Any category in .

[0024] The relevant credibility calculation process described in S102 and S103 is as follows:

[0025] Taking the credibility value A1 of the SAR image data in S102 as an example, the classification results are obtained after cross-training samples using the training set. The credibility calculation formulas for the classification results of different categories by different classifiers are as follows:

[0026]

[0027]

[0028]

[0029] Calculate the The probability that a class sample is correctly classified and necessity , and then calculate the classification credibility of this category .in, is the total number of samples to be tested, Indicates The number of class samples, Indicates Class samples are correctly divided into The number of classes, Indicates that other class samples are incorrectly classified as Calculate the classification credibility of each category each time during multiple cross-training , and find its average value A single model has different credibility for different categories of targets. The category credibility obtained by S102 , represents the classification credibility of the CNN model for samples of different categories in SAR data.

[0030] Calculate the classification accuracy of each cross-training , Represents the CNN model and finds its average value , the classification accuracy formula is as follows:

[0031]

[0032] Category credibility of S103 middle The probability that a class sample is classified correctly and necessity The calculation method is the same as above, except that Replace with , Replace with , represents the classification credibility of the MLP model for different categories of samples in the HRRP data, denoted as A2. At the same time, the corresponding , Represents the MLP model.

[0033] After the test set is input, the predicted classification result is obtained. Each sample in the test set There are corresponding prediction classification results And the posterior probability distribution ,in Divided into and , Represents classification into classes The results of the previous training set are used to find the corresponding and , denoted as A1 and A2 respectively.

[0034] The specific steps of constructing the initial confidence function in S104 are as follows:

[0035] (1) Use the information entropy formula to calculate the sample Corresponding and The entropy value and , the calculation formula is as follows:

[0036]

[0037] (2) Based on the calculated entropy value, determine the weight ratio of this part in the subsequent fusion, that is, the trust discount The larger the entropy value, the greater the posterior probability of the signal source, the greater the uncertainty, and the lower the assigned weight. The judgment formula is as follows:

[0038]

[0039] (3) Belief function redistribution. The specific implementation steps are as follows:

[0040] Computing the ensemble using fuzzy metrics , , the formula is as follows

[0041]

[0042] Redistribute the confidence function, the formula is as follows:

[0043]

[0044] in, and The product of ensures that the weight of evidence is proportional to its confidence during the fusion process. The correction factor added uses the complement of the original confidence function and the complement of the trust discount By multiplying by a smaller coefficient of 0.1, while reducing the reliance on low-confidence evidence, the information of such evidence is appropriately retained, avoiding over-reliance on a single source of evidence, strengthening the role of high-confidence evidence in decision-making, and providing a certain expression space for low-confidence evidence, thus achieving a more comprehensive and robust evidence integration, thereby improving the overall decision-making quality.

[0045] The normalization steps are shown in the following formula to achieve The sum of the data is 1.

[0046]

[0047]

[0048] The specific steps of S106 and S105 are as follows:

[0049] Each element in corresponds to a focal element in the evidence fusion rule, and they are all single focal elements (such as {A}), not compound focal elements (such as {A,B,C} or {A,C}). Bring it into the evidence theory fusion formula as the initial confidence function and calculate the confidence value after fusion , the specific implementation is as follows:

[0050]

[0051]

[0052] The largest median Corresponding category This is the final predicted classification result of the sample.

[0053] Directly extract the fused credibility value , recorded as A3, the formula is as follows:

[0054]

[0055] There is also a credibility value calculated from entropy uncertainty , recorded as A4, the specific calculation formula is as follows:

[0056]

[0057] The final classification credibility calculation process is shown below. The AA (Arithmetic Average) method is used to fuse the four credibility values ​​A1, A2, A3, and A4, and the fusion weight is selected using the Max-Min Optimization principle, in order to achieve a roughly equal contribution of the four credibility values ​​to the final credibility, avoiding any one of them having too much influence on the final result.

[0058] The calculation process is as follows:

[0059] (6a) Define KL divergence: For a sample Each credibility , first calculate its distribution , The format is A two-dimensional vector (with As an example, the distributions of the other three credibility levels are obtained in this way). Then calculate their distributions Relative to the fusion distribution The KL divergence of is as follows:

[0060]

[0061] (6b) Initialize fusion distribution: Assume that the initial fusion distribution =[0.5,0.5], initial weight , , , Both are 0.25.

[0062] (6c) Define the optimization problem: we want to find the weights , , , , so that the sum of KL divergences of all credibility is minimized.

[0063] (6d) Solve the optimization problem: Use the following formula:

[0064]

[0065] in .

[0066] The numerical optimization method is selected as SLSQP algorithm.

[0067] (6e) Update fusion distribution: After obtaining the current optimal weights, use these weights to update the fusion distribution .

[0068] (6f) Iterative optimization: Repeat (6a) to (6e) until the weight , , , converges to a stable value.

[0069] and According to and The target category corresponding to the maximum value Given, maximum The corresponding target category is the sample The predicted classification results are based on Calculate , . , , and That is, it corresponds to the four credibility levels A1, A2, A3, and A4 in the above steps. Then the sample The final confidence value of the classification result is , the calculation formula is as follows

[0070]

[0071] Ultimately, it is possible to obtain the credibility value of the prediction while obtaining the fusion prediction classification result.

[0072] The present invention has the following beneficial effects:

[0073] 1. The present invention uses a sampling extraction method to preprocess the HRRP data, compresses the data while retaining the original order and fluctuation characteristics, and improves the classification accuracy of the deep learning model when there are many target categories.

[0074] 2. The present invention uses the posterior probability distribution of the deep learning model to replace the requirement for the initial belief function in the evidence theory, avoiding the situation in which it is given by experts in the traditional method, making the initial belief function more objective.

[0075] 3. The present invention introduces the prior knowledge of the average classification accuracy of the training model and redistributes the weights of the initial confidence function in combination with entropy, thereby improving the reliability and robustness of the initial confidence function and the accuracy of ship classification.

[0076] 4. The present invention proposes a method for calculating the classification credibility of the classification model for different categories before fusion and the classification credibility of the sample after fusion, so that the credibility value of the result is output at the same time as the classification result, indicating the degree of recognition of the result. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of a ship classification method based on HRRP and SAR data trusted decision fusion provided by an embodiment;

[0078] Figure 2 is a visualization diagram of the HRRP preprocessing steps provided in the embodiment;

[0079] Figure 3 is a structural diagram of a SAR classification module provided in an embodiment;

[0080] Figure 4 It is a structural diagram of the HRRP classification module provided in the embodiment. DETAILED DESCRIPTION

[0081] The following further describes the embodiments and effects of the present invention in conjunction with the accompanying drawings. A ship classification method based on HRRP and SAR data trusted decision fusion is shown in FIG. Figure 1 As shown, the following steps are included:

[0082] Step 1: Build a dataset.

[0083] Simulate and obtain the HRRP one-dimensional data and SAR image data of the same ship target at the same time and the corresponding labels to obtain the original data set, and preprocess them separately. The total number of categories that need to be classified is , represent Any category in .

[0084] The SAR image is filtered by a Butterworth low-pass filter to enhance the contrast of the target area, making the features more obvious and convenient for training. The preprocessing steps can be expressed as follows:

[0085]

[0086]

[0087] in, is an image The pixel coordinates in is the filter kernel The element coordinates in is the distance between the midpoint of the frequency domain and the center of the frequency rectangle. The cutoff frequency is located at a distance from the origin. Where The transfer function of the 2nd-order Buttworth low-pass filter is , This is the image after filtering.

[0088] For HRRP data, such as Figure 2 As shown, the sampling feature extraction method is used to keep the original sequence characteristics and overall change rules as much as possible, which is specifically implemented as follows:

[0089]

[0090]

[0091]

[0092] in, represents the original HRRP data, Indicates the total number of distance units contained in the HRRP data. is the HRRP data output by preprocessing. The total number of distance units after processing is , It is a sampling extraction method.

[0093] The preprocessed data is divided into training set and test set in a ratio of 3:1. The training set will be used for training and evaluation of CNN and MLP models using cross-validation method, while the test set will be used to test the classification accuracy of the targets before and after fusion.

[0094] Step 2: Build and train a classification model for SAR images to obtain the required prior knowledge and credibility.

[0095] A CNN classification model is constructed to perform supervised training on the SAR image dataset. The training set is used for training, and the relevant credibility values ​​are recorded. The test set is predicted to give the posterior probability and classification prediction results of each sample being classified into a certain category.

[0096] The model consists of 5 convolutional layers, 3 pooling layers, 1 Flatten layer, 2 Dropout layers, and 3 fully connected layers. For the specific order, refer to Figure 3 .

[0097] The convolution kernel sizes of the five convolutional layers are 11×11, 5×5, 3×3, 3×3 and 3×3, respectively, with step sizes of 4, 1, 1, 1, 1; the number of convolution kernels is 96, 256, 384, 384, 256, respectively, and each convolutional layer is activated using ReLU; the three pooling layers use maximum pooling, and the kernel windows are 3×3, 2×2, 2×2, respectively, with step sizes of 4, 1, 1; the number of nodes in the three fully connected layers is 1024, 1024, N, respectively, and the number of output nodes in the last fully connected layer is Represents the number of categories that need to be classified, and its value is the posterior probability of the target multi-classification , representing The SAR data of samples is After (CNN model training) classification The probability of the category is , the category corresponding to the maximum posterior probability is the classification result. The training method is implemented as follows:

[0098] (2a) A portion of the training set samples is divided into a validation set with a training-validation ratio of 3:1, and cross-validation is started.

[0099] (2b) Initialize the convolutional neural network parameters.

[0100] (2c) The SAR of the input training sample set is forward propagated through the convolutional layer, pooling layer and fully connected layer of the convolutional neural network to obtain the output value.

[0101] (2d) The cross entropy loss function is used to calculate the error between the network output value and the target value.

[0102] (2e) Stochastic gradient descent SGD optimizer is used to update network parameters.

[0103] (2f) Repeat (2b) to (2e) until the error between the network output value and the target value is basically stable.

[0104] (2g) Input the validation set data, obtain the classification results, and calculate the classification accuracy of this training and classification credibility , the calculation formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] Calculate separately The probability that a class sample is classified correctly and necessity , and then calculate the classification credibility of this category .in, is the total number of samples to be tested, Indicates The number of class samples, Indicates Class samples are correctly divided into The number of classes, Indicates that other class samples are incorrectly classified as The number of classes.

[0110] (2h) Repeat (2a) to (2g) and calculate the average classification accuracy after multiple trainings , calculate the average credibility of the model for different categories after multiple trainings .

[0111] A single model has different credibility for different categories of targets. Category credibility obtained by S102 , represents the classification credibility of the CNN model for samples of different categories in SAR data.

[0112] After the test set is input, the predicted classification result is obtained. Each sample in the test set There are corresponding prediction classification results And the posterior probability distribution , find the corresponding one from the results of the previous training set , denoted as A1.

[0113] Step 3: Build and train the classification model of HRRP data to obtain the required prior knowledge and credibility, such as Figure 4 shown.

[0114] Construct an MLP classification model to perform supervised training on the HRRP dataset. Use the training set for training, record the relevant credibility values, and predict the test set to give the posterior probability of each sample being classified into a certain category and the classification prediction result.

[0115] The model consists of 1 input layer, 2 hidden layers, and 1 output layer.

[0116] Parameters: The number of neurons in the input layer is 2000, the number of neurons in the two hidden layers is 3000, 3000, and the number of neurons in the output layer is , represents the number of categories that need to be classified, and its value is the posterior probability of the target multi-classification , representing The HRRP data of samples After (MLP model training) classification The probability of the category is , the category corresponding to the maximum posterior probability is the classification result. The training method is implemented as follows:

[0117] (3a) A portion of the training set samples is divided into a validation set with a training-validation ratio of 3:1, and cross-validation is started.

[0118] (3b) Initialize the multilayer perceptron network parameters.

[0119] (3c) The HRRP of the input training sample set is forward propagated through the hidden layer (fully connected layer) of the MLP to obtain the output value.

[0120] (3d) The cross entropy loss function is used to calculate the error between the network output value and the target value.

[0121] (3e) Use the adam optimizer to update the network parameters.

[0122] (3f) Repeat (3b) to (3e) until the error between the network output value and the target value is basically stable.

[0123] (3g) Input the validation set data, obtain the classification results, and calculate the classification accuracy of this training and classification credibility , the calculation formula is as follows:

[0124]

[0125]

[0126]

[0127]

[0128] (3h) Repeat 3a) to 3g) and calculate the average classification accuracy after multiple trainings. , calculate the average credibility of the model for different categories after multiple trainings, and record it as , denoted as A2, Represents the number of categories that need to be classified.

[0129] A single model has different credibility for different categories of targets. Category credibility obtained by S103 , represents the classification confidence of the MLP model for samples of different categories in the HPPR data.

[0130] After the test set is input, the predicted classification result is obtained. Each sample in the test set There are corresponding prediction classification results And the posterior probability distribution , find the corresponding one from the results of the previous training set , denoted as A2.

[0131] Step 4: construct an initial confidence allocation module to redistribute the posterior probabilities obtained in S102 and S103, and construct two new sets of initial confidence functions derived from SAR data and HRRP data.

[0132] (4.1) Use the information entropy formula to calculate the sample Corresponding and The entropy value and , the calculation formula is as follows:

[0133]

[0134] (4.2) Based on the calculated entropy, determine the weight ratio of this part in the subsequent fusion, that is, the trust discount The larger the entropy value, the greater the posterior probability of the signal source, the greater the uncertainty, and the lower the assigned weight. The judgment formula is as follows:

[0135]

[0136] (4.3) Belief function redistribution, the specific implementation steps are as follows:

[0137] Computing the ensemble using fuzzy metrics , , the formula is as follows:

[0138]

[0139] Redistribute the confidence function, the formula is as follows:

[0140]

[0141] in, and The product of ensures that the weight of evidence is proportional to its confidence during the fusion process. The correction factor added uses the complement of the original confidence function and the complement of the trust discount By multiplying by a smaller coefficient of 0.1, while reducing the reliance on low-confidence evidence, the information of such evidence is appropriately retained, avoiding over-reliance on a single source of evidence, strengthening the role of high-confidence evidence in decision-making, and providing a certain expression space for low-confidence evidence, thus achieving a more comprehensive and robust evidence integration, thereby improving the overall decision-making quality.

[0142] The normalization steps are shown in the following formula to achieve The sum of the data is 1.

[0143]

[0144]

[0145] Step 5: Construct the evidence theory fusion module.

[0146] The two sets of initial confidence functions are combined to obtain the classification results of the target samples, and the confidence values ​​after the combination are recorded;

[0147] Each element in corresponds to a focal element in the evidence fusion rule, and they are all single focal elements (such as {A}), not compound focal elements (such as {A,B,C} or {A,C}). Substitute it into the evidence theory fusion formula to calculate the confidence value after fusion , the specific implementation is as follows:

[0148]

[0149]

[0150]

[0151] Directly extract the fused credibility value UA s , the formula is as follows:

[0152]

[0153] There is also a credibility value EA calculated from the entropy uncertainty s , the specific calculation formula is as follows:

[0154]

[0155] Step 6: Output the result and the credibility value.

[0156] The posterior probabilities saved in steps 2 and 3 are input into the initial confidence assignment module in step 4 and the fusion module in step 5 in turn to obtain the classification results of the fused test set. The credibility value of the fused sample classification is calculated and AA fusion calculation is performed on the previously recorded category credibility value to output the credibility corresponding to the classification result.

[0157] The maximum output after fusion The corresponding target category is the sample The predicted classification results.

[0158] The final classification credibility calculation process is shown below. The AA method is used to fuse the four credibility values, and the fusion weight is selected using the maximum and minimum optimization principle, in order to achieve a roughly equal contribution of the four credibility values ​​to the final credibility, avoiding any one of them having an excessive impact on the final result.

[0159] The calculation process is as follows:

[0160] (6a) Define KL divergence: For a sample Each credibility , first calculate its distribution , The format is A two-dimensional vector (with As an example, the distributions of the other three credibility levels are obtained in this way). Then calculate their distributions Relative to the fusion distribution The KL divergence of is as follows:

[0161]

[0162] (6b) Initialize fusion distribution: Assume that the initial fusion distribution =[0.5,0.5], initial weight , , , Both are 0.25.

[0163] (6c) Define the optimization problem: we want to find the weights , , , , so that the sum of KL divergences of all credibility is minimized.

[0164] (6d) Solve the optimization problem: Use the following formula:

[0165]

[0166] in .

[0167] The numerical optimization method is selected as SLSQP algorithm.

[0168] (6e) Update fusion distribution: After obtaining the current optimal weights, use these weights to update the fusion distribution .

[0169] (6f) Iterative optimization: Repeat (6a) to (6e) until the weight , , , converges to a stable value.

[0170] and According to and The target category corresponding to the maximum value Given, maximum The corresponding target category is the sample The predicted classification results are based on Calculate , . , , and That is, it corresponds to the four credibility levels A1, A2, A3, and A4 in the above steps. Then the sample The final confidence value of the classification result is , the calculation formula is as follows:

[0171]

[0172] Ultimately, it is possible to obtain the credibility value of the prediction while obtaining the fusion prediction classification result.

[0173] The effect of the present invention can be further illustrated by the following experiments.

[0174] 1. Experimental conditions:

[0175] The present invention conducts experiments on a simulation data set, which includes: 18 groups of SAR image data sets representing 18 categories, each group of 168 data sets are numbered, png format, and the size is 512*512, 18 groups of HRRP one-dimensional data sets representing 18 categories, each group of 168 data sets are numbered, csv format, the length varies, but all are greater than 2000, and correspond to the data in the SAR data set according to the number. A total of 3024*2 data sets.

[0176] The experimental hardware platform is Intel(R) Core(TM) i7-8550U CPU@1.80GHz 2.00 GHz CPU, 8GBRAM;

[0177] The experimental software platform is Python 3.6.

[0178] 2. Experimental content and results analysis:

[0179] Based on the model training method of the present invention, a training set and a test set are divided. The training set is trained to obtain the CNN and MLP classification models and the average recognition accuracy of the models. The test set is used for test comparison and input into the two models to obtain the classification results and posterior probability values. The improved DS method is used to fuse the posterior probabilities to obtain the final classification results. The classification results are shown in Table 1 below. The evaluation indicators are precision, recall and F1 score.

[0180] Table 1 Comparison of classification accuracy before and after fusion

[0181]

[0182] Compared with the classification effect of a single sensor, the fusion classification model of the present invention is superior to the classification result of a single sensor in terms of precision, recall rate and F1 score indicators, which verifies the superiority of the present invention.

[0183] While outputting the classification results, the present invention will also give a credibility value for each recognition result for each sample. For example, in the experiment, the classification result of sample 1 is the 8th category, and the credibility value is 97.35%, which is relatively high. Its label is the 8th category, and the classification is correct; the classification result of sample 2 is the 5th category, and the credibility value is 67.27%, which is relatively low. Its label is the 4th category, and the classification is wrong, which proves the effective role of the credibility value.

[0184] In summary, the ship classification method based on the fusion of HRRP and SAR data trustworthy decision-making proposed in the present invention improves the accuracy of SAR and HRRP fusion classification, and gives the credibility value corresponding to the classification result, which has practical application value.

Claims

1. A ship classification method based on HRRP and SAR data trusted decision fusion, characterized by: The following steps are involved: S101: obtaining HRRP one-dimensional data and SAR image data of the same ship at the same time and corresponding labels, preprocessing them respectively, and dividing them into a training set and a test set; S102: Construct a CNN classification model and an MLP classification model, perform supervised training on the SAR image data and the HRRP data using the training set, respectively, obtain the credibility values ​​A1 and A2, and predict the test set to give the posterior probability of each sample being classified into a certain category and the classification prediction result; The calculation process of the credibility value is as follows: For the credibility value A1 obtained from the SAR image data, the classification results are obtained by cross-training samples using the training set. The credibility calculation formula for the classification results of different categories by different classifiers is as follows: SA(n)=(P pos (n)+P nec (n)) / 2 Calculate the probability P of the nth class sample being correctly classified pos (n) and necessity P nec (n), and then calculate the classification credibility of this category SA(n), S is the total number of samples to be tested, S n Indicates the number of samples in the nth category, ST n Indicates the number of samples of the nth class that are correctly classified into n classes, SF n Indicates the number of other class samples that are incorrectly classified into n classes; Calculate the classification confidence SA(n) of each category in each cross-training and find its average value A single model has different credibility for different categories of targets, and the obtained category credibility Represents the classification confidence of the CNN model for samples of different categories in SAR data; Calculate the classification accuracy acc(E1) of each cross-training, where E1 represents the CNN model, and find its average value The classification accuracy formula is as follows: For the category credibility of HRRP data, the probability P of n-class samples being classified correctly in HA(n) is pos (n) and necessity P nec The calculation method of (n) is consistent with that of SAR image data, and the classification confidence HA(n) is obtained, and the average value Represents the classification credibility of the MLP model for different categories of samples in the HRRP data, and calculates the corresponding E2 represents the MLP model; After the test set is input, the predicted classification result is obtained. Each sample s in the test set has a corresponding predicted classification result z i (i=1,2,…,N) and the posterior probability distribution P s (z i |E j ), where E j Divided into E1 and E2, i Represents the result classified as class n, find the corresponding one from the results of the training set and Denoted as A1 and A2; S103: constructing an initial confidence allocation module, redistributing the two posterior probabilities obtained in S102, and constructing two new sets of initial confidence functions derived from SAR data and HRRP data; The specific process of constructing the initial confidence function is as follows: S103.

1. Use the information entropy formula to calculate the P corresponding to sample s s (z i |E1) and P s (z i |E2) entropy value H s (E1) and H s (E2), the calculation formula is as follows: S103.

2. According to the entropy value, determine the weight ratio of this part in the subsequent fusion, that is, the trust discount T s : S103.

3. Computing the full set P using fuzzy metrics s (z0|E j ), z0=Θ={1,2,3...N}: Redistribute the belief function: reP s (z i |E j )=P s (z i |E j )·T s (E j )+0.1·(1-P s (z i |E j ))·(1-T s (E j )),i=0,1,2…N,j=1,2 The normalization steps are shown below to achieve bpa s (Z|E j )Z=[z0,z1,z2,…,z N ]The sum of the data is 1: BPA s (Z|E j )={reP s (z0|E j ) / σ(E j ),reP s (z2|E j ) / σ(E j ),…,reP s (z. N |E j ) / σ(E j )},j=1,2; S104: construct an evidence theory fusion module, fuse the two sets of initial confidence functions to obtain the classification results of the target samples, and record the fused confidence values ​​A3 and A4; S105: Perform AA fusion calculation on the credibility values ​​A3 and A4 and the credibility values ​​A1 and A2, and output the credibility corresponding to the classification result.

2. The ship classification method based on HRRP and SAR data trusted decision fusion according to claim 1 is characterized in that: The preprocessing is specifically implemented as follows: The SAR image is filtered by a Butterworth low-pass filter to enhance the contrast of the target area: X′(i,j)=∑ u ∑ v B(u,v)·X(i-u,j-v) Where (i, j) is the pixel coordinate in the image X, (u, v) is the element coordinate in the filter kernel B, D(u, v) is the distance between the midpoint in the frequency domain and the center of the frequency rectangle, the transfer function of the d-order Butterworth low-pass filter with a cutoff frequency at a distance from the origin D0 is B(u, v), and X′ is the image after filtering; For HRRP data, sampling feature extraction method is used for preprocessing: H W =f c (H) x i =h |i·M / W| i=1,2,3,…W-1 x W =h M Where H = [h1,h2,…,h M ] represents the original HRRP data, M represents the total number of distance units contained in the HRRP data, and H w =[x1,x2,…,x W ] is the HRRP data output by preprocessing. The total number of distance units after processing is W, f c (·) is the sampling extraction method.

3. The ship classification method based on HRRP and SAR data trusted decision fusion according to claim 2 is characterized in that: The step S104 is specifically implemented as follows: bpa s (Z|E j ) corresponds to a focal element in the evidence fusion rule, and they are all single focal elements. s (Z|E j ) is introduced into the evidence theory fusion formula as the initial confidence function, and the confidence value P after fusion is calculated s (z i |DS): P s P with the largest median value of (Z|DS) s (z i |DS) is the final predicted classification result of the sample; Directly extract the fused credibility value UA s , denoted as A3: DO s =1-P s (z0|DS) The confidence value EA calculated from the entropy uncertainty s , recorded as A4:

4. The ship classification method based on HRRP and SAR data trusted decision fusion according to claim 3 is characterized in that: The specific implementation process of performing AA fusion calculation and outputting the credibility corresponding to the classification result is as follows: S105.

1. Define KL divergence: For each credibility a of sample s, first calculate its distribution f a (x), f a (x) format is The distribution of the other three credibility levels is obtained in this way, and then their distribution f is calculated. a (x) relative to the fusion distribution f AA KL divergence of (x): S105.

2. Initialize fusion distribution: Assume that the initial fusion distribution f AA (x) = [0.5, 0.5], the initial weights w1, w2, w3, w4 are all 0.25; S105.3, define the optimization problem: we want to find the weights w1, w2, w3, w4, w1+w2+w3+w4=1, which minimizes the sum of KL divergences of all credibility; S105.

4. Solve the optimization problem: Use the following formula: where f AA (x)=w1f1(x)+w2f2(x)+w3f3(x)+w4f4(x) Select the numerical optimization method as SLSQP algorithm; S105.

5. Update the fusion distribution f using the current optimal weight AA (x); S105.

6. Repeat S105.1 to S105.5 until weights w1, w2, w3, w4 converge to stable values; S s (n) and HA s (n) According to P s (z i |E1) and P s (z i |E2) the target category n corresponding to the maximum value is given, the maximum P s (z i |DS) is the predicted classification result of sample s. s (z i |DS) calculates UA s ,EA s ; Then the final credibility value of the classification result of sample s is FinA s , the calculation formula is as follows: Ultimately, it is possible to obtain the credibility value of the prediction while obtaining the fusion prediction classification result.