Synthetic Aperture Radar Target Trust Recognition Method Based on Evidence Deep Learning

Abstract

The present invention provides a method for synthetic aperture radar target trust recognition based on evidence deep learning, which includes obtaining multiple SAR images as training samples; constructing an evidence deep learning trust recognition model with a feature representation module, a confidence assignment module, and an expected utility module connected in sequence; in the training stage, only considering the exact behaviors where the possible recognition behaviors are the same as the actual categories, using the training samples to perform iterative training on it by adjusting the weights and biases through backpropagation until the possible behaviors are the same as the actual categories; in the subsequent application stage, since the possible recognition behaviors in the expected utility module are all non-empty subsets of the actual categories, selecting the behavior with the maximum expected utility value as the predicted label output of the target image. The present invention realizes trust recognition by compromising between recognition accuracy and imprecision, thereby improving the recognition accuracy and the reliability of the recognition results of SAR image targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and particularly relates to a method for synthetic aperture radar target trust recognition based on evidence deep learning. Background Art

[0002] Synthetic aperture radar has the characteristics of strong penetration and strong weather adaptability, and has been widely used in many military and civilian fields. With the improvement of SAR sensor performance, the requirements for SAR data processing are gradually increasing, and SAR automatic recognition technology is also increasingly put into use or in the preparation stage. In recent years, deep learning has developed rapidly, and a large number of new network model architectures have been proposed, which can be applied to SAR automatic recognition technology.

[0003] Traditionally, for recognition and classification problems, when learning a deep learning model, the input space is divided into as many decision regions as there are classes. The model can predict the possible probability values of each class of the input sample, and select the class with the largest probability value as the final recognition result according to the predicted probability to achieve accurate recognition. However, in the case of high uncertainty, this hard division of the input space usually leads to misrecognition. For example, observations near the boundary of the decision region are ambiguous, where multiple classes have similar probabilities, and the model prediction result has great unreliability at this time. Rejection is a classic method to solve this problem. By quantifying the uncertainty of the prediction and avoiding making a decision when the uncertainty is too high, the occurrence of misrecognition can be reduced. However, although rejection can effectively improve the recognition accuracy, this completely rejecting method usually cannot provide enough effective information for decision-making. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method for synthetic aperture radar target trust recognition based on evidence deep learning. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0005] A method for synthetic aperture radar target trust recognition based on evidence deep learning provided by the present invention includes:

[0006] Step 1: Obtain multiple SAR images as training samples;

[0007] Step 2: Construct an evidence deep learning trust recognition model;

[0008] Among them, the constructed evidence deep learning trust recognition model includes a feature representation module, a confidence assignment module, and an expected utility module connected in sequence;

[0009] Step 3: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence, so that the feature representation module extracts the features of the training samples, the confidence assignment module performs confidence assignment based on the extracted features, and the expected utility module calculates the expected utility values of possible recognition behaviors, and adjusts the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model until the training is completed;

[0010] Step 4: Input the target image to be recognized into the evidence deep learning trust recognition model at the end of training, so that the evidence deep learning trust recognition model performs feature extraction, confidence assignment, and calculation of the expected utility values of possible recognition behaviors on the test samples, and selects the behavior with the maximum expected utility as the predicted recognition behavior output of the target image.

[0011] Optionally, the construction of the evidence deep learning trust recognition model includes:

[0012] Step 2-1: Obtain a basic convolutional neural network model composed of a convolutional layer, a pooling layer, and a fully connected layer;

[0013] Step 2-2: Set the convolution kernels of the convolutional layer, the pooling sizes of the pooling layer, and the number of categories of the fully connected layer to obtain a feature representation module;

[0014] Step 2-3: According to the evidence theory, construct a confidence assignment module, which is used to calculate multiple groups of basic trust assignments of each category with respect to the class center for the category set, aggregate the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregate the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the category power set, forming a mass vector;

[0015] Step 2-4: Construct an expected utility module for

[0016] Define the behavior set and the original utility matrix for the category set;

[0017] Among them, each behavior in the behavior set represents a specific behavior of classifying a sample into a certain category, and each general term in the original utility matrix represents the utility value of belonging the sample of the true category to other categories;

[0018] Generalize the behavior set to assign each sample to any non-empty subset of the category set;

[0019] Make a decision using the generalized behavior set to expand the original utility matrix into an extended utility matrix;

[0020] Among them, the utility value of the exact classification of each item in any non-empty subset of the general items in the extended utility matrix is obtained by ordered weighted average aggregation;

[0021] Define the imprecise tolerance of the general items in the extended utility matrix;

[0022] Taking the imprecise tolerance as a constraint condition, calculate the utility value of the ordered weighted average aggregation operation by maximizing the cross entropy;

[0023] Based on the extended utility matrix and the mass vector, use the generalized Hurwicz criterion to calculate the expected utility when classifying the sample into any non-empty subset;

[0024] Determine the recognition behavior with the maximum expected utility value of any non-empty subset as the true category of the sample.

[0025] Optionally, in step 2-3, the basic belief assignment is:

[0026]

[0027] The Euclidean distance between the input x and the i-th (i = 1,..., n) class center of the j-th class is:

[0028]

[0029] j = 1,..., M, i = 2,..., n, then the mass function of the j-th class obtained through the class center is

[0030]

[0031] The mass of all non-empty elements in the class power set is:

[0032]

[0033] Among them, and is related to The parameter, the class set Ω = {ω1,..., ω M}, its power set is 2 Ω The input x is represented by a P-dimensional feature vector. For each class j (j = 1,..., M), the convolutional neural network can learn to obtain n class centers, denoted as m 1,s (·) is the combined result of m 1 (·),..., m s (·), and m 1,1 (·) = m 1 (·), then the mass values of all non-empty elements in the class power set are obtained, forming the mass vector m.

[0034] Optionally, the set of actions in steps 2-4 is represented as fωi represents classifying the sample into class ω i , and the original utility matrix is U M×M , and the general term u ij ∈[0,1] represents classifying the sample with the true class ω j into ω i when the utility value. The generalized set of actions is f A represents classifying into any non-empty subset A, and the extended utility matrix is whose general term represents the utility when classifying the sample with the true class ω j into any non-empty subset A;

[0035] The general term value of the extended utility matrix is obtained by aggregating the utility values of each precise classification in any non-empty subset A through ordered weighted average OWA. The aggregation method is

[0036]

[0037] where is the set composed of the elements in the original utility matrix U M×M , and the k-th largest element in . The weight g=(g1,...,g|A|) represents the preference for the choice when the classifier needs to make a precise decision among a set of possible choices. The elements in the weight vector g represent the tolerance for imprecision in the decision;

[0038] The tolerance for imprecision of the general term in the extended utility matrix is:

[0039]

[0040] The cross entropy is represented as

[0041]

[0042] where TDI(g)=γ is the constraint condition, Σgk = 1 and gk≥0;

[0043] The expected utility when classifying the sample into any non-empty subset A is calculated using the generalized Hurwicz criterion as:

[0044]

[0045] where the pessimistic index v is a model hyperparameter, representing the attitude of the decision towards fuzziness, E m (f A ) and respectively represent the lower and upper bounds of the expected utility, and the calculation methods are as follows

[0046]

[0047]

[0048] The final recognition and classification result A of the sample is the category with the maximum expected utility value:

[0049]

[0050] Optionally, step 3 includes:

[0051] Step 3-1: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence;

[0052] Step 3-2: The feature representation module extracts the features of the training samples and inputs them into the confidence assignment module;

[0053] Step 3-3: The confidence assignment module calculates multiple groups of basic trust assignments of each training sample with respect to the class center for each category according to the extracted features, aggregates the basic trust assignments of the same category, obtains the combined mass functions of multiple categories, and aggregates the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the category power set;

[0054] Step 3-4: The mass values of all non-empty elements in the category power set are formed into a mass vector of the training sample and input into the expected utility module;

[0055] Step 3-5: The expected utility module calculates the expected utility of attributing the training sample to any non-empty subset of the set of all categories according to the extended utility matrix;

[0056] Step 3-6: The recognition behavior with the maximum expected utility value in any non-empty subset is determined as the true category of the training sample;

[0057] Step 3-7: Adjust the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model;

[0058] Step 3-8: Repeat steps 3-1 to 3-8 until the training cut-off condition is reached, and obtain the evidence deep learning trust recognition model at the end of training.

[0059] Optionally, the training cut-off condition in step 3-8 includes:

[0060] The evidence deep learning trust recognition model is trained to reach the preset number of training times;

[0061] The error of the evidence deep learning trust recognition model is less than the error threshold or no longer changes;

[0062] The test accuracy of the test deep learning trust recognition model meets the accuracy requirements.

[0063] Optionally, step 4 includes:

[0064] Step 4-1: Input the target image to be recognized into the evidence deep learning trust recognition model that has stopped training;

[0065] Step 4-2: The feature representation module extracts the features of the target image and inputs them into the belief assignment module;

[0066] Step 4-3: The belief assignment module calculates multiple groups of basic belief assignments of the target image with respect to the class center for each category according to the extracted features of each target image, aggregates the basic belief assignments of the same category to obtain the combined mass functions of multiple categories, and aggregates the basic belief assignments of different categories to obtain the mass values of all non-empty elements in the power set of categories;

[0067] Step 4-4: The mass values of all non-empty elements in the power set of categories are formed into a mass vector of the target image and input into the expected utility module;

[0068] Step 4-5: The expected utility module calculates the expected utility of belonging the target image to any non-empty subset of the set of all categories according to the extended utility matrix;

[0069] Step 4-6: Determine the recognition behavior with the maximum expected utility value in any non-empty subset as the true category of the target image.

[0070] Advantages of the present invention:

[0071] The present invention provides a method for synthetic aperture radar target trust recognition based on evidence deep learning. By obtaining multiple SAR images as training samples; constructing an evidence deep learning trust recognition model with a feature representation module, a belief assignment module, and an expected utility module connected in sequence; only considering the exact behavior where the possible recognition behavior is the same as the actual category during the training stage, and using the training samples to perform iterative training on it by adjusting the weights and biases through backpropagation until the possible behavior is the same as the actual category; in the subsequent application stage, since the possible recognition behaviors in the expected utility module are all non-empty subsets of the actual category, select the behavior with the maximum expected utility value as the predicted label output of the target image. The present invention realizes trust recognition by making a compromise between recognition accuracy and imprecision, thereby improving the recognition accuracy and reliability of the SAR image target.

[0072] The following will further elaborate on the present invention in conjunction with the drawings and embodiments. Description of the Drawings

[0073] Figure 1 is a schematic flowchart of a synthetic aperture radar target trust recognition method based on evidence deep learning provided by an embodiment of the present invention;

[0074] Figure 2a is an overall architecture diagram of an evidence deep learning trust recognition model provided by an embodiment of the present invention;

[0075] Figure 2b is a general CNN network architecture diagram provided by an embodiment of the present invention;

[0076] Figure 3 is a diagram showing examples of 10 - class sample pictures of the MSTAR dataset;

[0077] Figure 4 is a diagram showing an example of the recognition result of the EDL - CR model;

[0078] Figure 5 is a schematic diagram showing the change of the average utility with respect to v. Detailed Embodiments

[0079] The following further describes the present invention in detail with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0080] Before introducing the specific solution of the present invention, first explain the influence of sample classification.

[0081] Compared with complete rejection, classifying samples into subsets of classes is more reasonable. This set - valued recognition (or credal recognition) method can better utilize the advantages of deep learning in extracting and representing data features, provide more effective information for decision - making, and at the same time can minimize the risk of misrecognition and improve the reliability of recognition results.

[0082] As Figure 1 shown, a synthetic aperture radar target trust recognition method based on evidence deep learning provided by the present invention includes:

[0083] Step 1: Obtain multiple SAR images as training samples;

[0084] It should be noted that: in actual operation of the present invention, multiple SAR images can be obtained from the radar for annotation and then used as training samples.

[0085] Step 2: Construct an evidence deep learning trust recognition model;

[0086] Among them, the constructed evidence deep learning trust recognition model includes a feature representation module, a confidence assignment module, and an expected utility module that are connected in sequence;

[0087] Step 3: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence, so that the feature representation module extracts the features of the training samples, the confidence assignment module performs confidence assignment based on the extracted features, and the expected utility module calculates the expected utility value of possible recognition behaviors, and adjusts the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model until the training is terminated;

[0088] Step 4: Input the target image to be recognized into the evidence deep learning trust recognition model at the end of training, so that the evidence deep learning trust recognition model performs feature extraction, confidence assignment, and calculation of the expected utility value of possible recognition behaviors on the test sample, and selects the behavior with the maximum expected utility as the predicted recognition behavior output of the target image.

[0089] The present invention provides a method for synthetic aperture radar target trust recognition based on evidence deep learning. By obtaining multiple SAR images as training samples; constructing an evidence deep learning trust recognition model including a feature representation module, a confidence assignment module, and an expected utility module that are connected in sequence; only considering the exact behaviors where the possible recognition behaviors are the same as the actual categories during the training phase, and using the training samples to perform iterative training on it by adjusting the weights and biases through backpropagation until the possible behaviors are the same as the actual categories; in the subsequent application phase, since the possible recognition behaviors in the expected utility module are all non-empty subsets of the actual categories, select the behavior with the maximum expected utility value as the predicted label output of the target image. The present invention realizes trust recognition by making a compromise between recognition accuracy and imprecision, thereby improving the recognition accuracy of SAR image targets and the reliability of recognition results.

[0090] Embodiment 2

[0091] As an optional embodiment of the present invention, the construction of the evidence deep learning trust recognition model includes:

[0092] Step 2-1: Obtain a basic convolutional neural network model composed of a convolutional layer, a pooling layer, and a fully connected layer;

[0093] Step 2-2: Set the convolutional kernel of the convolutional layer, the pooling size of the pooling layer, and the number of categories of the fully connected layer to obtain a feature representation module;

[0094] Feature representation is an important part of deep learning work, which includes discovering predictors for recognition and classification from raw data. In recent years, due to the ability to construct rich deep feature representations, deep learning models have achieved excellent performance in tasks such as pattern recognition and segmentation. Convolutional neural network (CNN) is a special multi-layer neural network and one of the most widely used deep learning architectures. It is a deep feed-forward network very suitable for processing images. It can reduce the dimensionality of large amounts of image data into small amounts while effectively retaining image features, and has the characteristics of local connection and weight sharing.

[0095] The most common CNN mainly includes convolutional layers, pooling layers and fully connected layers. The network architecture is shown in Figure 2. The convolutional layers and pooling layers convert the input data into an intermediate representation and act as feature extractors. Generally speaking, a deep CNN consists of several stacked convolutional layers and pooling layers, which can process raw data and convert it into a higher-level feature map. Then, the fully connected layer acts as a decision maker and assigns the input to one of the classes according to the feature map. Therefore, the final output at the stacked level in the CNN can be regarded as the feature representation of the input data. In the evidence deep learning trust recognition model, these high-level features extracted and represented by the network will be used as the input of the confidence assignment module.

[0096] Based on the classic neural network model LetNet-5, the stacked level composed of its convolutional layers and pooling layers in the present invention is used as a feature extraction and representation module. The size of the convolutional kernel used in the convolutional layer is 5×5, and the maximum pooling method is adopted in the pooling layer, and the pooling size is selected as 2×2. The network structure settings of the LeNet-5 model are shown in Table 1, where K is equal to the number of categories in the dataset used.

[0097] Table 1 LeNet-5 Network Structure Settings

[0098]

[0099] Step 2-3: According to the evidence theory, construct a confidence assignment module, which is used to calculate multiple groups of basic trust assignments of each category with respect to the class center for the category set, aggregate the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregate the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the category power set, and form a mass vector;

[0100] As an extension of traditional probability, evidence theory provides a general framework for uncertainty modeling and reasoning. It represents uncertainty and other information through concepts such as belief functions. Evidence theory is built on a frame of discernment, which represents a non-empty set of all possible states of a problem, denoted as Ω = {ω1,..., ω M}, for any result, it can be represented by a subset of Ω. The set consisting of all possible subsets of Ω is called the power set of Ω, denoted as 2 Ω . The basic belief assignment function (also known as the mass function) is a mapping from 2 Ω to [0, 1], denoted as m: 2 Ω → [0, 1]. m(A) represents the basic belief assignment function corresponding to proposition A, which represents the degree of support of the evidence for proposition A and satisfies:

[0101]

[0102] where indicates that the confidence of the impossible event is 0, indicates that the sum of the confidences of all elements in 2 Ω is 1. If A satisfies m(A) > 0, then A is called a focal element under the frame of discernment. It should be noted that m(A) only represents the confidence value assigned to A itself and does not include the confidence values assigned to its proper subsets. The confidence value assigned to the entire frame of discernment Ω is called the global uncertainty, denoted as m(Ω).

[0103] Two mass functions m1 and m2 based on the same frame of discernment Ω and representing independent evidence items can be combined together through the Dempster's rule. The fusion process is as follows:

[0104]

[0105] where

[0106]

[0107] The conflict factor k quantifies the conflict between the mass functions m1 and m2. Only when the two evidences are not completely conflicting, i.e., k < 1, the Dempster's combination rule is valid. The Dempster's combination rule satisfies the commutative law and the associative law. Therefore, it is easy to generalize Equation (2) to the fusion of multiple evidences.

[0108] Consider an M-class recognition problem, the class set Ω = {ω1,..., ω M}, its power set is 2 Ω , and the input x is represented by a P-dimensional feature vector. For each class j (j = 1,..., M), the neural network can learn to obtain n class centers, denoted as In the neural network, the class center can be regarded as the network connection weight. The Euclidean distance between the input x and the i-th (i = 1,..., n) class center of class j is

[0109]

[0110] For the input x, the closer its distance to the class center is, the greater the possibility that it belongs to this class, and the corresponding basic belief assignment value should also be higher. The basic belief assignment based on distance is

[0111]

[0112] where and are parameters related to . For each class, n sets of basic belief assignments regarding the class center can be obtained and aggregated through the Dempster combination rule. The combined mass function is obtained through iteration, that is

[0113]

[0114] where j = 1,..., M, i = 2,..., n, then the final mass function of the j-th class obtained through the class center is

[0115]

[0116] Suppose the set-valued class elements with a cardinality greater than or equal to 2 in the power set 2 Ω form a set Ψ. Then, for the M sets of mass values obtained after combination, the BCR combination rule is used for inter-class fusion, that is

[0117]

[0118] where m 1,s (·) is the combination result of m 1 (·),..., m s (·), and m 1,1 (·) = m 1 (·). After fusion, the mass values of all non-empty elements in the class power set are obtained, forming a mass vector m, which is used as the input of the expected utility module to calculate the expected utility values of all possible classification behaviors finally.

[0119] Step 2 - 4: Construct an expected utility module for

[0120] Step a: For the class set, define the behavior set of this class set and the original utility matrix;

[0121] where each behavior in the behavior set represents a specific behavior of classifying a sample into a certain class, and each general term in the original utility matrix represents the utility value of classifying a sample of the true class into other classes;

[0122] Step b: Generalize the behavior set to assigning each sample to any non-empty subset of the class set;

[0123] Step c: Making a decision using the generalized set of actions to expand the original utility matrix into an extended utility matrix;

[0124] Among them, the utility values of the precise classification of each item in any non-empty subset of the general terms in the extended utility matrix are aggregated through ordered weighted averaging;

[0125] It should be noted that: Let Ω = {ω1,..., ω M} be the set of categories. For an identification classification problem with only precise prediction, an action (act) is defined as assigning a sample to one and only one of the M categories. The set of actions is where f ωi represents classifying the sample into category ω i . To make a decision, a utility matrix U M×M is defined, where the general term u ij ∈[0,1] represents the utility value when a sample with the true category ω j is classified as ω i . When making a decision, each action f ωi will generate an expected utility, and U M×M is called the original utility matrix. For an identification classification problem with imprecise prediction, the action is generalized to assign a sample to any non-empty subset A of Ω, and then the set of actions becomes where f A represents classification into the subset A. To make a decision using the generalized set of actions, the original utility matrix U M×M is expanded into an extended utility matrix whose general term represents the utility when a sample with the true category ω j is classified as A.

[0126] The general term value of the extended utility matrix is obtained by aggregating the utility values of the precise classification of each item in A through ordered weighted averaging (OWA), and the aggregation method is

[0127]

[0128] where is the set M×M composed of the elements in the original utility matrix U and is the k-th largest element in it, and the weight g = (g1,..., g|A|) represents the preference for the choice when the classifier needs to make a precise decision among a set of possible choices. The elements in the weight vector g represent the tolerance of the decision to imprecision. For example, if for any imprecise set A, as long as it contains the true category, the utility value of the action f A is 1, then at this time, it is completely tolerant of imprecision. In this case, only the set The maximum utility value, i.e., (g1,..., g|A|) = (1, 0,..., 0).

[0129] Step d: Define the imprecision tolerance of the general terms in the extended utility matrix;

[0130] Step e: Using the imprecision tolerance as a constraint condition, calculate the utility value of the ordered weighted average aggregation operation by maximizing the cross - entropy;

[0131] It should be noted that: The method for determining the weight vector g is as follows. First, define the imprecision tolerance as

[0132]

[0133] Its maximum value is 1, the minimum value is 0, and the average value is 0.5. In fact, we only need to consider γ values between 0.5 and 1 because when γ is less than 0.5, more precise classification is preferred. Given an imprecision tolerance γ, the weights of the OWA operation can be calculated by maximizing the cross - entropy, and the cross - entropy is expressed as

[0134]

[0135] It satisfies the constraint conditions TDI(g) = γ, Σgk = 1 and gk ≥ 0. Table 2 shows the extended utility matrix obtained by the OWA operation with γ = 0.7 for a three - classification problem. The first three rows form the original utility matrix, indicating that the utility value is 1 when the sample is classified into its true class, and 0 otherwise. The remaining rows are the utility values generated by aggregation. For example, for a sample with the true class ω1, the utility when classifying it as {ω1, ω2} is 0.7.

[0136] Step f: Based on the extended utility matrix and the mass vector, use the generalized Hurwicz criterion to calculate the expected utility when classifying the sample into any non - empty subset;

[0137] Step g: Determine the true class of the sample as the recognition behavior with the maximum expected utility value of any non - empty subset.

[0138] Based on the extended utility matrix and the output m of the belief assignment module, use the generalized Hurwicz criterion to calculate the expected utility when classifying the sample into A, that is

[0139]

[0140] where the pessimistic index v is a model hyper - parameter, representing the decision - making attitude towards fuzziness. E m (f A ) and respectively represent the lower and upper bounds of the expected utility, and the calculation methods are respectively

[0141]

[0142]

[0143] The final recognition and classification result A of the sample is the category with the maximum expected utility value, that is

[0144]

[0145] Table 2 Extended utility matrix when γ = 0.7

[0146]

[0147] Embodiment III

[0148] As an alternative embodiment of the present invention, step 3 includes:

[0149] Step 3-1: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence;

[0150] Step 3-2: The feature representation module extracts the features of the training samples and inputs them into the confidence assignment module;

[0151] Step 3-3: The confidence assignment module calculates multiple groups of basic trust assignments of each training sample with respect to the class center for each category according to the extracted features, aggregates the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregates the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the category power set;

[0152] Step 3-4: The mass values of all non-empty elements in the category power set are formed into the mass vector of the training sample and input into the expected utility module;

[0153] Step 3-5: The expected utility module calculates the expected utility of attributing the training sample to any non-empty subset of the set of all categories according to the extended utility matrix;

[0154] Step 3-6: The recognition behavior with the maximum expected utility value in any non-empty subset is determined as the true category of the training sample;

[0155] Step 3-7: Adjust the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model;

[0156] Step 3-8: Repeat steps 3-1 to 3-8 until the training cut-off condition is reached, and obtain the evidence deep learning trust recognition model at the end of training;

[0157] Among them, the training termination conditions include:

[0158] Training the deep learning trust recognition model for evidence reaches a preset number of training times;

[0159] The error of the deep learning trust recognition model for evidence is less than the error threshold or no longer changes;

[0160] The test accuracy of the deep learning trust recognition model for evidence meets the accuracy requirements.

[0161] Embodiment 4

[0162] As an optional embodiment of the present invention, step 4 includes:

[0163] Step 4-1: Input the target image to be recognized into the deep learning trust recognition model for evidence with training termination;

[0164] Step 4-2: The feature representation module extracts the features of the target image and inputs them into the confidence assignment module;

[0165] Step 4-3: The confidence assignment module calculates multiple groups of basic trust assignments of the target image with respect to the class center for each category according to the extracted features of each target image, aggregates the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregates the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the category power set;

[0166] Step 4-4: The mass values of all non-empty elements in the category power set are formed into a mass vector of the target image and input into the expected utility module;

[0167] Step 4-5: The expected utility module calculates the expected utility of attributing the target image to any non-empty subset of the set of all categories according to the extended utility matrix;

[0168] Step 4-6: Determine the true category of the target image as the recognition behavior with the maximum expected utility value in any non-empty subset.

[0169] Next, the technical effects of the present invention are verified through experiments. The present invention is evaluated and verified on the MSTAR dataset.

[0170] (1) Preprocessing of the experimental dataset

[0171] The MSTAR dataset is one of the most commonly used datasets in current SAR image target recognition research. Classified according to different acquisition conditions, the MSTAR dataset can be divided into two categories: Extended Operating Conditions (EOC) and Standard Operating Conditions (SOC). In SOC, the SAR images for testing and training have the same target shape configuration and model, and only the pitch angle and azimuth angle of the target are different during imaging. In this experiment, the data collected under the SOC acquisition conditions are used, which include a total of 10 types of target images. The training set data is collected at a pitch angle of 17°, and the test set data is collected at a pitch angle of 15°. The training set contains a total of 2,747 sample data, and the test set contains a total of 2,425 sample data. The detailed information of the dataset is shown in Table 3. Since the initial data sizes are inconsistent, the image data is first preprocessed, and the images are uniformly scaled to a size of 64×64 pixels, and each type of sample picture in the training set and test set is labeled respectively. Figure 3 shows examples of 10 types of sample pictures and their corresponding optical images.

[0172] Table 3 Data Information of the MSTAR Dataset under SOC Conditions

[0173]

[0174] (2) Experimental Evaluation Metrics

[0175] In this experiment, the recognition results of the evidence deep learning trust recognition model are first evaluated through the average expected utility and average cardinality. The larger the average expected utility value and the smaller the average cardinality, the better the recognition result.

[0176] For dataset T, the average expected utility is

[0177]

[0178] where y i is the true class of sample i, and A(i) is the final recognition result of sample i, represents the utility value when classifying i as A. When only considering the exact classification behavior, the AU criterion defined in Equation (16) is the exact recognition accuracy.

[0179] The calculation method of the average cardinality of the recognition result is

[0180]

[0181] When evaluating using the recognition accuracy, for the trust recognition result, it is considered a correct recognition when the true class is included in the final result. Let T represent the situation where the true class is included in the recognition result, and F represent the situation where the true class is not included in the recognition result. Then the recognition accuracy is

[0182]

[0183] (3) Recognition performance comparison experiment

[0184] To verify the target recognition performance of the proposed EDL-CR model of the present invention, this method was compared with the baseline method LeNet-5 under the same experimental conditions. Figure 4 First, the recognition effect of EDL-CR is shown, where 0, 6, 7, and 9 represent classes 2S1, BTR-70, T62, and ZIL-131 respectively.

[0185] Before evaluating the trust recognition performance of the EDL-CR model, first, for a given uncertainty tolerance value γ, its corresponding optimal pessimistic index v is determined. Figure 5 The change in the average utility of the recognition results when the v value changes under different γ values is shown. It can be seen from the figure that when γ is greater than 0.7, the change in the v value has a relatively large impact on the average utility. The optimal v value is selected for the experiment according to the results in the figure.

[0186] Table 4 and Table 5 respectively show the changes in the average utility and average cardinality of the recognition results with respect to the imprecision tolerance γ under the optimal pessimistic index v value. It can be seen from the results that as the γ value increases, both the average utility and the average cardinality gradually increase, and when γ = 1, that is, when completely tolerating imprecision, the average utility of the recognition results is equal to 1.0, the average cardinality is the same as the number of classes, and at this time all samples are classified as the universal set Ω. This shows that the trust recognition effect of the model is greatly affected by the imprecision tolerance value, and the larger the γ value, the more inclined the model is to perform imprecise recognition.

[0187] Table 4 Average utility under different imprecision tolerances γ

[0188]

[0189] Table 5 Average cardinality under different imprecision tolerances γ

[0190]

[0191] When considering the recognition accuracy of the model, Table 6 shows the changes in the recognition accuracy of EDL-CR under different γ values.

[0192] Table 6 Recognition accuracy under different imprecision tolerances γ

[0193]

[0194] Considering both the accuracy and the imprecision, γ = 0.8 is selected as the model parameter for the SAR image target recognition task of the MSTAR dataset. Table 7 shows the comparison between the EDL-CR and the LeNet-5 baseline methods in terms of accuracy, average cardinality, and average utility metrics under this parameter.

[0195] Table 7 Comparison of Method Performance

[0196]

[0197] As can be seen from Table 7, the EDL-CR model proposed in the present invention makes a compromise between accuracy and imprecision. When selecting an appropriate γ value, compared with the baseline method, although the degree of imprecision increases, the recognition accuracy has been greatly improved with a limited increase in the average cardinality.

[0198] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0199] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases.

[0200] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for identifying the trust of synthetic aperture radar targets based on evidence deep learning, characterized in that, Including: Step 1: Obtain multiple SAR images as training samples; Step 2: Construct an evidence deep learning trust recognition model; Among them, the constructed evidence deep learning trust recognition model includes a feature representation module, a belief assignment module, and an expected utility module connected in sequence; Step 3: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence, so that the feature representation module extracts the features of the training samples, the belief assignment module performs belief assignment according to the extracted features, and the expected utility module calculates the expected utility value of possible recognition behaviors, and adjusts the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model until the training stops; Step 4: Input the target image to be recognized into the evidence deep learning trust recognition model at the end of training, so that the evidence deep learning trust recognition model performs feature extraction, belief assignment, and calculation of the expected utility value of possible recognition behaviors on the test sample, and selects the behavior with the maximum expected utility as the predicted recognition behavior output of the target image; The construction of the evidence deep learning trust recognition model includes: Step 2-1: Obtain a basic convolutional neural network model composed of a convolutional layer, a pooling layer, and a fully connected layer; Step 2-2: Set the convolution kernel of the convolutional layer, the pooling size of the pooling layer, and the number of categories of the fully connected layer to obtain a feature representation module; Step 2-3: According to the evidence theory, construct a belief assignment module, which is used to calculate multiple groups of basic belief assignments of each category with respect to the class center for the category set, aggregate the basic belief assignments of the same category to obtain the combined mass functions of multiple categories, and aggregate the basic belief assignments of different categories to obtain the mass values of all non-empty elements in the category power set, forming a mass vector; Step 2-4: Construct an expected utility module for For the category set, define the behavior set and the original utility matrix of this category set; Among them, each behavior in the behavior set represents a specific behavior of classifying a sample into a certain category, and each general term in the original utility matrix represents the utility value of attributing a sample of the true category to other categories; Generalize the behavior set to assign each sample to any non-empty subset of the category set; Make a decision using the generalized behavior set so that the original utility matrix is extended to an extended utility matrix; Among them, the utility value of each item accurately classified in any non-empty subset of the general term in the extended utility matrix is aggregated through ordered weighted averaging; Define the imprecision tolerance of the general term in the extended utility matrix; Taking the imprecision tolerance as a constraint condition, calculate the utility value of the ordered weighted averaging aggregation operation by maximizing the cross entropy; Based on the extended utility matrix and the mass vector, use the generalized Hurwicz criterion to calculate the expected utility when classifying a sample into any non-empty subset; Determine the true category of the sample as the recognition behavior with the maximum expected utility value of any non-empty subset; In Step 2-3, the basic belief assignment is: The Euclidean distance between the input x and the i-th (i = 1,..., n) class center of category j is: For \(j = 1,\cdots,M\) and \(i = 2,\cdots,n\), the mass function obtained by the \(j\)-th class through the class center is The mass of all non-empty elements in the power set of the category is: Among them, and are parameters related to . The category set Ω = {ω1,..., ω M}, and its power set is 2 Ω . The input x is represented by a P-dimensional feature vector. For each category j (j = 1,..., M), the convolutional neural network can learn to obtain n class centers, denoted as m 1,s (·) is the combined result of m 1 (·),..., m s (·), and m 1,1 (·) = m 1 (·). Then, the mass values of all non-empty elements in the category power set are obtained, forming the mass vector m; The behavior set in Step 2-4 is represented as fωi represents classifying the sample into class ω i , and the original utility matrix is U M×M , and the general term u ij ∈[0,1] represents the utility value when the true class is ω j and the sample is classified into ω i . The generalized behavior set is f A represents classifying into any non-empty subset A, and the extended utility matrix is whose general term represents the utility when the true class is ω j and the sample is classified into any non-empty subset A; The general term value of the extended utility matrix is obtained by aggregating the utility values of each precise classification in any non-empty subset \(A\) through ordered weighted average OWA, and the aggregation method is Among them, is a set composed of elements in the original utility matrix U M×M , and the k-th largest element in . The weight g = (g1,..., g|A|) represents the preference for the choice when the classifier needs to make an accurate decision among a set of possible choices. The elements in the weight vector g represent the tolerance of the decision to imprecision; The imprecise tolerance of the general term in the extended utility matrix is: The cross entropy is expressed as where \(TDI(g)=\gamma\) is the constraint condition, \(\sum g_k = 1\) and \(g_k\geq0\); Using the generalized Hurwicz criterion, the expected utility when classifying a sample into any non-empty subset \(A\) is calculated as: Among them, the pessimistic index v is a model hyperparameter, representing the decision-making attitude towards fuzziness. E m (f A ) and represent the lower bound and upper bound of the expected utility respectively, and the calculation methods are as follows The final recognition classification result \(A\) of the sample is the category with the maximum expected utility value:

2. The synthetic aperture radar target trust recognition method based on evidence deep learning according to claim 1, characterized in that The said step 3 includes: Step 3-1: Input the training samples into the pre-constructed evidence deep learning trust recognition model in sequence; Step 3-2: The feature representation module extracts the features of the training samples and inputs them into the confidence assignment module; Step 3-3: The confidence assignment module calculates, according to the features of each training sample extracted, multiple groups of basic trust assignments of the training sample with respect to the class center for each category, aggregates the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregates the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the power set of the category; Step 3-4: Input the mass values of all non-empty elements in the power set of the category, which form the mass vector of the training sample, into the expected utility module; Step 3-5: The expected utility module calculates the expected utility of attributing the training sample to any non-empty subset of all category sets according to the extended utility matrix; Step 3-6: Determine the true category of the training sample as the recognition behavior with the maximum expected utility value in any non-empty subset; Step 3-7: Adjust the weights and biases in the evidence deep learning trust recognition model through backpropagation to minimize the error of the evidence deep learning trust recognition model; Step 3-8: Repeat steps 3-1 to 3-8 until the training cut-off condition is reached to obtain the evidence deep learning trust recognition model at the end of training.

3. The method for synthetic aperture radar target trust recognition based on evidence deep learning according to claim 2, characterized in that, The training cut-off conditions in step 3-8 include: The training of the evidence deep learning trust recognition model reaches the preset number of training times; The error of the evidence deep learning trust recognition model is less than the error threshold or no longer changes; The test accuracy of the test deep learning trust recognition model meets the accuracy requirements.

4. A method for synthetic aperture radar target trust recognition based on evidence deep learning according to claim 1, characterized in that, The said step 4 includes: Step 4-1: Input the target image to be recognized into the evidence deep learning trust recognition model at the end of training; Step 4-2: The feature representation module extracts the features of the target image and inputs them into the confidence assignment module; Step 4-3: The confidence assignment module calculates, according to the features of each target image extracted, multiple groups of basic trust assignments of the target image with respect to the class center for each category, aggregates the basic trust assignments of the same category to obtain the combined mass functions of multiple categories, and aggregates the basic trust assignments of different categories to obtain the mass values of all non-empty elements in the power set of the category; Step 4-4: Input the mass values of all non-empty elements in the power set of the category, which form the mass vector of the target image, into the expected utility module; Step 4-5: The expected utility module calculates the expected utility of attributing the target image to any non-empty subset of the set of all categories according to the extended utility matrix; Step 4-6: Identify the recognition behavior with the maximum expected utility value in any non-empty subset as the true category of the target image.

Images