Robust SAR image recognition system and method integrating feature loss and three-objective optimization

Abstract

The present invention discloses a robust synthetic aperture radar image recognition system and method that integrates feature loss and three-objective optimization. A multi-feature weighted loss function based on multiple hidden layers of a neural network model is constructed. The neural network architecture parameters of the SAR image recognition model, the weight parameters of the multi-feature weighted loss function, and the training learning rate are individually encoded. The number of parameters of the SAR image recognition model, the recognition accuracy of normal samples, and the recognition accuracy of adversarial samples are used as optimization targets. An offline optimization platform based on the three-objective optimization method is designed, and a SAR image recognition model with lightweight, high precision, and strong adversarial robustness for online deployment is obtained. The technology of the present invention not only realizes the automatic generation of a SAR image recognition model that takes into account multiple performance indicators, but also the constructed multi-feature weighted loss function expands the feature combination method of the SAR image recognition model, and also improves the comprehensive performance of the model, such as lightweight, recognition accuracy, and adversarial robustness.

Description

Technical Field

[0001] The present invention relates to an automatic deep learning technology for the information security of Synthetic Aperture Radar (SAR) image recognition, which achieves optimal comprehensive performance, including lightweight models, high recognition accuracy, and strong robustness to adversarial attacks. In particular, the present invention relates to a robust SAR image recognition system and method that integrates feature loss and three-objective optimization. Background Art

[0002] Because Synthetic Aperture Radar (SAR) systems are unaffected by weather and can provide high-resolution surveillance images of target areas at any time, they have been widely used in battlefield surveillance, fire control, guidance, navigation, resource surveying, mapping, ocean surveillance, and environmental remote sensing. In these areas involving important national strategic deployments, the recognition accuracy and robustness of SAR image recognition models are particularly important. Recognition accuracy is directly related to the identification of the target, while model robustness is related to the model's ability to resist external malicious attacks. Therefore, in response to the increasing demand for SAR application deployment and national strategic security requirements, the development of lightweight SAR image detection systems with high recognition accuracy, strong robustness, and low resource consumption is urgently needed.

[0003] Current SAR image recognition models are highly integrated with deep learning techniques, which has improved their accuracy for normal samples to a certain extent. However, this has also exposed them to adversarial attack techniques from the deep learning field. Many adversarial attack methods (such as FSGM, PGD, CW, Deepfool, Jitter, and Random Noise) can effectively reduce the recognition accuracy of classical image recognition neural network models, demonstrating the poor robustness of existing neural network architectures and their vulnerability to adversarial attacks from malicious adversaries. Current neural network architectures for SAR image recognition models are still based on classical image recognition models and are hand-designed based on expert experience. These architectures suffer from complex structures, poor robustness, and difficulty in flexibly switching between different recognition scenarios. Automated design of SAR image recognition model architectures can effectively address these issues, but the inherent speckle noise in SAR images increases the difficulty of model training, making automated machine learning techniques used in traditional computer vision difficult to directly apply to SAR image recognition models. Searching for lightweight SAR image recognition model architectures that take into account the properties of SAR images, with high recognition accuracy and robustness to adversarial attacks, is a pressing technical challenge in the field of SAR image recognition.

[0004] In existing technical research, the adversarial attack defense technology for SAR image recognition models only includes SAR image adversarial sample detection methods based on multi-target optimal feature attribution selection and SAR adversarial image detection systems and methods based on multi-target neural architecture optimization. However, the former is mainly used to identify whether the input image is an adversarial sample, without considering the adversarial attack robustness of the SAR image recognition model itself. After the adversarial sample detection, the image still needs to be input into the SAR image recognition model for recognition, which consumes additional computing resources and communication transmission resources. The latter has the following technical deficiencies: (1) This technology is essentially an automatic architecture design method for the SAR image recognition model. In addition to the model architecture and parameters, it does not consider the organic integration and optimization design of multiple deep and shallow feature loss functions of the SAR image during the model training process, resulting in the model recognition accuracy still having room for improvement; (2) This technology randomly selects an adversarial attack to test the adversarial sample recognition accuracy during the fitness evaluation stage, without considering the strength of the adversarial attack. Some individuals use adversarial samples with weak attack power (such as FGSM) for testing, while some individuals use adversarial samples with strong attack power (such as PGD) for testing, which affects the optimization direction of the adversarial attack robustness during the optimization process. However, the development of a technology that comprehensively uses adversarial samples with strong attack power for fitness evaluation and simultaneously considers the architecture parameters and feature loss parameters of the SAR image recognition model for parallel optimization has not been reported. Therefore, in response to the engineering requirements of SAR image recognition application scenarios for optimal comprehensive performance such as model lightweight, high recognition accuracy and strong adversarial robustness, automatic deep learning technology that integrates model neural architecture search, hyperparameter optimization and multi-feature loss function optimization is urgently needed to be developed. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the existing technology and provide a robust SAR image recognition system and method that integrates feature loss and three-objective optimization.

[0006] The object of the present invention is achieved through the following technical solutions: a robust SAR image recognition system integrating feature loss and three-objective optimization, characterized in that the system includes a SAR image data preprocessing module, an adversarial robust SAR image recognition model offline training and optimization module integrating multi-feature loss and three-objective optimization, and a SAR image online recognition module;

[0007] The SAR image data preprocessing module collects image data of the SAR system's real-time monitoring from the SAR system's real-time database. During this process, there is a threat of adversarial attacks from malicious adversaries, resulting in the presence of adversarial samples in the obtained real-time monitoring image data. The real-time monitoring image data is subjected to image data normalization and image data normalization processing to generate an online detection dataset, and the online detection dataset is transmitted to the SAR image online recognition module. The SAR image data preprocessing module collects historical image data of the SAR system from the SAR system's historical database, and after image data normalization and image data normalization processing, generates an offline dataset. The offline dataset is divided into an offline training set and an offline validation set in a ratio of 4:1, and then transmitted to an offline training and optimization module of an adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization.

[0008] The offline training and optimization module of the adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization first performs integer encoding on the neural network architecture parameters of the SAR image recognition model, and performs real number encoding on multiple loss weight parameters and training learning rates, which are collectively used as individual mixed encodings; then, an initialized candidate population is randomly generated, and the offline training set is trained based on the SAR image recognition model architecture corresponding to different individual encodings, the training learning rate, and the multi-feature fusion loss obtained by weighted calculation of multi-feature loss of cross entropy loss, similar shallow loss, similar deep loss, heterogeneous shallow loss, and heterogeneous deep loss; the trained SAR image recognition model is used to calculate the recognition accuracy of normal samples of the offline validation set and the number of model parameters, and the SAR image adversarial samples are generated through the PGD adversarial attack module. The adversarial sample recognition accuracy of the model is calculated; the normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters of the SAR image recognition model are used as fitness indicators of the individual three-objective fitness function evaluation submodule, and the fast non-dominated sorting, crowding distance calculation and fitness comprehensive evaluation submodules are used to sort the individuals in the initial candidate population, generate the parent population according to the population size, and select the optimal individual from the obtained Pareto front individuals; the parent population is input into the hybrid differential mutation module and the hybrid differential crossover module, and after the offspring population is generated, the parent population and the offspring population are merged to produce a new generation of candidate population; the above evolutionary process is repeated until the maximum number of evolutionary rounds is reached, thereby obtaining the final optimal individual; the optimal SAR image recognition model corresponding to the optimal individual is transmitted to the SAR image online recognition module;

[0009] The SAR image online recognition module deploys the optimal SAR image recognition model online, performs image recognition on the online detection dataset, and calculates the normal sample recognition accuracy. It then generates adversarial samples of the online detection dataset through the Fast Gradient Sign (FGSM) attack module, the Gradient Iteration (PGD) attack module, the Optimization (CW) attack module, the Decision Boundary (Deepfool) attack module, the Jitter (Jitter) attack module, and the Random Noise (Random Noise) attack module. Using the optimal SAR image recognition model deployed online, the model's classification accuracy for the above adversarial samples is calculated, thereby obtaining the model's adversarial sample recognition accuracy.

[0010] The robust SAR image recognition method integrating feature loss and three-objective optimization comprises the following steps:

[0011] (1) The SAR image data preprocessing module obtains the SAR image data stored in the historical monitoring process from the SAR system as the original data set, and marks it as O; the image data normalization module is used to perform image data normalization operation on O to obtain the normalized data set X; the image data normalization module is composed of a segmentation operation and a scaling operation; the segmentation operation divides the SAR image in the data set into regions. When the input SAR image contains multiple recognition targets, it is segmented into multiple SAR images so that each SAR image contains only a single recognition target. When the input SAR image contains only a single recognition target, no operation is performed; the scaling operation adjusts the size of the SAR image in the data set to keep it consistent with the internal parameters of the SAR image recognition model; the image data normalization operation is performed on the normalized data set X using formula (1) to obtain the offline data set X o ; Then, according to the ratio of 4:1, o Divide into offline training set X o_train and offline validation set X o_val ;

[0012]

[0013] (2) Setting the relevant parameter values ​​in the offline training and optimization module of the robust SAR image recognition model that integrates multi-feature loss and three-objective optimization, the parameter values ​​include the population size N, the number of layers of the SAR image recognition model architecture combination Layers, the number of units in the SAR image recognition model architecture combination layer c, and the lower limit of the normal architecture unit parameter Normal low , the upper limit of the normal architecture unit parameter Normal up , constraining the lower limit of the architecture unit parameters Reduction low , constraining the upper limit of the architecture unit parameters Reduction up, the lower limit of the residual architecture control parameter Restype low , the upper limit of the residual architecture control parameter Restype up , the lower limit of the loss weight parameter Loss low , the upper limit of the loss weight parameter Loss up , the lower limit of the learning rate parameter lr low , the upper limit of the learning rate parameter lr up , training round EP train , mutation scaling factor σ, crossover rate β, maximum evolutionary round E max ;

[0014] (3) Encode the architecture parameters as integers, and encode the loss weights and training learning rates as real numbers, which are used together as individual mixed encodings in the three-objective optimization method. Initialize 2N individuals as the initial candidate population Q h , where the encoding form of each individual is Indi = [Normal Unit, Reduction Unit, Restype, Loss Weights, lr], Indi represents any individual in the population, where Normal Unit, Reduction Unit and Restype together constitute the architecture parameters of the SAR image recognition model, Loss Weights represents the loss weight parameter of the SAR image recognition model, and lr represents the training learning rate of the SAR image recognition model;

[0015] The specific encoding forms of Normal Unit and Reduction Unit are the same, both are [(o1 1 ,n1 1 ,o2 1 ,n2 1 ),(o1 2 ,n1 2 ,o2 2 ,n2 2 ),...,(o1 c ,n1 c ,o2 c ,n2 c )],remember(o1 i ,n1 i ,o2 i ,n2 i ) is the i-th subunit, where 1≤i≤c; for the i-th subunit, n1 i Represents the first operation node in the subunit, o1 i Represents the operation node n1 i The specific operation, n2 i Represents the second operation node in the unit, o2i Represents the operation node n2 i Specific operation; j i is the specific operation on the jth node in the i-th unit, where 1≤j≤2, o j i The corresponding relationship with the specific operation is: when o j i = 0, perform the maximum pooling operation; when o j i =1, perform average pooling operation; when o j i = 2, no operation is performed; when o j i = 3, a depth-separable convolution operation with a convolution kernel size of 3×3 is performed; when o j i = 4, a depth-separable convolution operation with a convolution kernel size of 5×5 is performed; when o j i = 5, a dilated convolution operation with a kernel size of 3×3 is performed; when o j i = 6, a dilated convolution operation with a convolution kernel size of 5×5 is performed; when o j i =7, a dilated convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 8, a depth-wise separable convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 9, the spatially separable convolution operations with kernel sizes of 1×3 and 3×1 are performed; when o j i = 10, the spatially separable convolution operations with kernel sizes of 1×5 and 5×1 are performed; when o j i = 11, the spatially separable convolution operations with kernel sizes of 1×7 and 7×1 are performed respectively; for Normal Unit and Reduction Unit, the difference between the two is that when o j i When coming from Normal Unit, j i The parameters involved in the step length in the corresponding specific operations are all set to 1; when o j i When coming from Reduction Unit, o j i The parameters involving step length in the corresponding specific operations are all set to 2;

[0016] The correspondence between Normal Unit encoding and Reduction Unit encoding and model architecture is as follows: States is the set of intermediate feature values ​​of the model architecture corresponding to the current Unit encoding, and its initial value is States = [s1, s2], s1, s2 are the intermediate features of the model transferred to the current Unit; traverse the current Unit encoding in turn, according to n1 i and n2 i The value of selects the intermediate feature of the corresponding position in States, using o1 i and o2 i The corresponding operation processes the intermediate feature values ​​and merges the results to obtain the new model intermediate features and add them to States. Finally, the intermediate features that have not been selected in States except s1 and s2 are merged as the output of the model architecture corresponding to the current Unit encoding;

[0017] The specific form of Restype encoding is [Res1, Res2, ..., Res Layers ], used to control whether the model architecture corresponding to each Unit encoding adopts residual connection. k The model architecture corresponding to the k-th Unit encoding, where 1≤k≤Layers; when Res k = 0, the model architecture corresponding to the Unit encoding does not use residual connection, and the corresponding mode of the model architecture does not change; when Res k =1, the model architecture corresponding to the Unit encoding adopts residual connection, and the output of the Unit is merged with the initial feature s2 input to the Unit as the new output;

[0018] The correspondence between individual architecture codes and SAR image recognition model architectures is as follows: Based on the number of SAR image recognition model architecture combination layers (Layers) and the residual control code (Restype), the model architecture corresponding to the Normal Unit code and the model architecture corresponding to the Reduction Unit code are sequentially selected for concatenation. The model architecture corresponding to each Unit code is considered a layer of model architecture combination layer. After cross-stacking in the order of Normal Unit, Reduction Unit, Normal Unit, ReductionUnit, ..., Normal Unit, Reduction Unit, a neural network model with the number of layers is formed. A compression excitation network module is added after each layer of model architecture combination layer, and a global pooling layer is added after the last layer of model architecture combination layer.

[0019] The specific form of Loss Weights encoding is [w1,w2,w3,w4,w5]; for the SAR image recognition model architecture corresponding to an individual, w1 is the cross entropy loss weight, w2 is the same shallow loss weight, w3 is the same deep loss weight, w4 is the different shallow loss weight, and w5 is the different deep loss weight, and their values ​​are all greater than 0;

[0020] The specific form of lr encoding is [lr]. For the SAR image recognition model architecture corresponding to an individual, lr is the learning rate of the model during offline training, and its value is greater than 0;

[0021] Initial candidate population Q h The specific generation process is as follows:

[0022] (3.1) According to the lower limit of the normal architecture unit parameter Normal low , the upper limit of the normal architecture unit parameter Normal up , constraining the lower limit of the architecture unit parameters Reduction low , constraining the upper limit of the architecture unit parameters Reduction up , the lower limit of the residual architecture control parameter Restype low , the upper limit of the residual architecture control parameter Restype up , the lower limit of the loss weight parameter Loss low , the upper limit of the loss weight parameter Loss up , the lower limit of the learning rate parameter lr low , the upper limit of the learning rate parameter lr up , generate N initial individuals Indis, the integer code value of the initial individual is a random integer between the lower and upper limits of the corresponding parameter, and the real number code value is a random real number between the lower and upper limits of the corresponding parameter;

[0023] (3.2) Note Indi up =[Normal up ,Reduction up ,Restype up ,Loss up ,lr up ], Indi low =[Normal low ,Reduction low ,Restype low ,Loss low ,lr low ]; the opposite individual of individual Indi in the coding space is recorded as Indi con ,Indi conThe definition of is shown in formula (2); Generate N opposite individuals Indis con ;

[0024] Indi con =Indi low +Indi up -Indi (2)

[0025] (3.3) N initial individuals Indis and N opposite individuals Indis con A total of 2N individuals serve as the initial candidate population Q h ;

[0026] (4) Label the current evolutionary round of the multi-objective optimization technique as E, and set E = 0;

[0027] (5) For the candidate population Q h The offline training and optimization module of the adversarial robust SAR image recognition model integrating multi-feature loss and three-objective optimization is based on Q h The model architecture parameters corresponding to the individuals in the neural network model are used to generate the neural network model, and then the weighted loss of the model is calculated according to the Loss Weights parameter as the total loss of the model l sum , using the lr parameter as the training learning rate of the model, and using the adaptive moment estimation optimizer (Adam) to optimize the offline training set X o_train The model is trained with SAR image data in EP. train ;

[0028] The total loss of the model is l sum Use formula (3) to calculate, where l ce represents the cross entropy loss of the model, Represents the same shallow loss of the model, represents the same deep loss of the model, represents the heterogeneous shallow loss of the model, Represents the heterogeneous deep loss of the model. ce , Calculate according to formula (4) to formula (8) respectively;

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] Where θ represents all the parameters that constitute the current model, x represents the image data sample in the offline training set, and y represents the category corresponding to x. One-hot encoding is used. The implementation process of one-hot encoding is as follows;

[0036] (5.1) The correct category of the image is y true ,y n is the number of total classification categories;

[0037] (5.2) Generate a line of length y n The all-zero vector y one_hot ;

[0038] (5.3) one_hot Middle y true The value at the position is changed to 1;

[0039] (5.4)y one_hot That is category y true One-hot encoding of ;

[0040] Where n represents the total number of classification categories of samples in the offline training set, y (m) represents the mth value of y, f(x) (m) represents the m-th bit value in the prediction result of the current model for the image data sample x, that is, the prediction probability of the m-th class; p m Represents the number of image data samples of the mth category in the current batch training data, and Represents two different image data samples belonging to the mth category in the current batch training data, where f m ≠g m , and Represents the fth n Class and g n Two image data samples of the class, where f n ≠g n Where f surface (x) represents the shallow feature output of the current model for the image data sample x, that is, the feature value output by the first hidden layer of the current model, and the feature value is subjected to maximum pooling processing; f deep(x) represents the deep feature output of the current model for the image data sample x, that is, the eigenvalue output by the penultimate hidden layer of the current model, and the eigenvalue is subjected to maximum pooling processing; A(a,b) is the permutation number calculation, which represents the total number of permutation schemes for sequentially selecting b elements from a elements for permutation, where a and b are both positive integers, and the calculation process is shown in formula (9);

[0041]

[0042] After training, use the offline validation set X according to formula (10) o_val Calculate the normal sample recognition accuracy of the SAR image recognition model Based on the iterative gradient adversarial attack (PGD), the adversarial sample set of the offline verification set is generated, and the adversarial sample recognition accuracy of the model is calculated according to formula (11): where Num clean_right Represents the number of samples correctly identified in the offline validation set, Num adv_right Represents the number of correctly identified samples in the adversarial sample set of the offline validation set, Num all Represents the total number of samples in the offline validation set; counts the number of internal parameters of the SAR image recognition model Params;

[0043]

[0044]

[0045] (6) For the candidate population Q h The individuals in the population are subjected to fast non-dominated sorting, crowding sorting and fitness comprehensive evaluation operations. The negative number of the individual's normal sample recognition accuracy is recorded as f1, the negative number of the individual's adversarial sample recognition accuracy is recorded as f2, and the number of model parameters Params is recorded as f3. The fast non-dominated sorting and crowding sorting select N individuals as the parent population Q. The fitness comprehensive evaluation operation calculates the fitness comprehensive evaluation value of the Pareto frontier individuals obtained after the fast non-dominated sorting, and selects the optimal individual. Specifically, the fast non-dominated sorting measures the three evaluation indicators of the individual's normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters, and divides the individuals to obtain the initial candidate population Q. h Pareto frontier individuals;

[0046] When individual Indi i The three optimization objective function values ​​of are all smaller than those of another individual Indi j When the individual is called Indi i Dominant individual Indi j . Remember n i is the number of individuals that dominate the i-th individual in the population, Si is the set of individuals dominated by the i-th individual; the specific implementation process of fast non-dominated sorting is:

[0047] (6.1) Let F1 be the empty set;

[0048] (6.2) Traverse the population Q in sequence h Individuals in, calculate the n corresponding to each individual i value, and will be controlled by the individual

[0049] Other individuals are placed in the corresponding set S i middle;

[0050] (6.3) n i = 0 individuals are put into the set F1;

[0051] (6.4) Let j1 = 1;

[0052] (6.5) Let H be the empty set;

[0053] (6.6) Traversal All individuals in , denoted by S u For The set of individuals dominated by the u-th individual, n q To dominate S u The number of individuals in the qth individual; get Each individual in S u ;

[0054] (6.7) For all S u , traverse each S u For all individuals in , calculate the n corresponding to each individual q value, and let n q =n q -1;

[0055] (6.8) If n q =0, then S u The qth individual is put into the set H;

[0056] (6.9) Let j1 = j1 + 1;

[0057] (6.10)

[0058] (6.11) Repeat steps (6.5) to (6.10) until you get is an empty set;

[0059] After completing the fast non-dominated sort, the parent population Q of size N is generated by combining the congestion sort. The specific implementation process is as follows:

[0060] (6.12) Let Q be the empty set, j2 = 1;

[0061] (6.13) Calculation The crowding distance of all individuals in . i for The crowding distance of the i-th individual in is calculated by using f1, f2 and f3 respectively. Sort the individuals in the , and take the three individuals with the largest f1 value, the largest f2 value and the largest f3 value as the boundary individuals. The crowding degree d i The crowding distance of all individuals except the boundary individuals is calculated according to formula (12). z i+1 express f of the i+1th individual z Target fitness value, f z i-1 express f of the i-1th individual z Target fitness value;

[0062]

[0063] (6.14) Let |Q| be the number of individuals in Q, for The number of individuals in . When All individuals in are put into Q. When Individuals are sorted from large to small according to the crowding distance. Select N-|Q| individuals in turn and put them into Q;

[0064] (6.15) Let j2 = j2 + 1;

[0065] (6.16) Repeat steps (6.12) to (6.15) until |Q| = N;

[0066] After that, the fitness comprehensive evaluation operation is performed to select the best individual. The individuals in F1 are taken as Pareto frontier individuals, and the best individual is selected by combining the three optimization objective function values. The specific selection process is as follows: when F1 contains only one individual, the individual is selected as the best individual in this evaluation. best When F1 contains two or more individuals, the comprehensive fitness evaluation value f of the i-th individual in F1 is calculated according to formula (13): i all , where f1 i represents the optimization target value of f1 for the i-th individual, f2 irepresents the optimization target value of f2 for the i-th individual, f3 i represents the f3 optimization target value of the i-th individual. Select f i all The smallest individual is the optimal individual Indi best , when there are multiple individuals with the smallest f i all When selecting The individual with the higher value is regarded as the optimal individual of the model best ;

[0067]

[0068] (7) For each individual in the population Q, a hybrid differential mutation module is used to perform a mutation operation, and the mutated population is marked as Q m , the specific implementation process is:

[0069] (7.1) Let i1 = 0, Q m is an empty set, and the i1th individual in Q is Indi i1 ;

[0070] (7.2) To India i1 Randomly select three other individuals Indi r1 ,Indi r2 and Indi r3 , where i1≠r1≠r2≠r3;

[0071] (7.3) According to the fitness comprehensive evaluation operation in step (6), select Indi r1 ,Indi r2 and Indi r3 The best individual in rb ;

[0072] (7.4) i1 The mutant offspring individual is recorded as Indi muti1 Indi(Arch) represents the value of the integer code position in the individual Indi code, Indi(Other) represents the value of the real code position in the individual Indi code, Indi muti1 Integer encoding part Indi muti1 The variation value of (Arch) is calculated by formula (14), Indi muti1 (Other) The variation value of the real number coding part is calculated by formula (15), where σ represents the variation scaling factor, Represents a floor operation;

[0073]

[0074] Indi muti1 (Other)=Indi rb (Other)+σ×(Indi r2 (Other)-Indi r3 (Other)) (15)

[0075] (7.5)Indi muti1 (Arch) and Indi muti1 (Other) after splicing to get Indi muti1 , will Indi muti1 The values ​​exceeding the boundary value in the code are modified to the boundary value;

[0076] (7.6) Let i1 = i1 + 1;

[0077] (7.7) Repeat steps (7.2) to (7.6) until every individual in Q has undergone mutation. Put all the mutated offspring individuals into the population Q m middle;

[0078] (8) For population Q m For each individual in the , a hybrid difference crossover module is used for crossover operation, and the population after crossover is marked as Q x , the specific implementation process is:

[0079] (8.1) Let i2 = 0, Q x is an empty set, let Q m The i2th individual is Indi muti2 , the i2th individual in Q is Indi i2 ;

[0080] (8.2) Note Indi muti2 The crossover offspring individual is Indi ci2 ,Indi muti2 The encoding length is l Indi2 . Generate a 1 to l Indi2 A uniformly distributed random integer in the range k1, where 1≤k1≤l Indi2 , let k1 be the fixed cross position;

[0081] (8.3) Note Indi ci2 (k2) is Indi ci2 The value of the k2th position on the , Rand(k2) generates a uniformly distributed random number in the range of 0 to 1, where 0≤Rand(k2)≤1, and Indi is calculated according to formula (16) ci2 The value of each bit in ;

[0082]

[0083] (8.4) The individual Indi ci2 The values ​​outside the boundary value in the code are modified to the boundary value;

[0084] (8.5) Let i2 = i2 + 1;

[0085] (8.7) Repeat steps (8.2) to (8.5) until Q m Each individual in the crossover operation is performed. All the obtained crossover offspring individuals are placed into the population Q x middle;

[0086] (9) Combine population Q with population Q x Merge into a new candidate population Q hnew ;

[0087] (10) Let Q h =Q hnew , E=E+1;

[0088] (11) Repeat steps (5) to (10) until E = E max ;

[0089] (12) For the E max Candidate population Q for rounds of evolution h According to the steps (5) and (6), fast non-dominated sorting and fitness comprehensive evaluation are performed to obtain the optimal individual Indi best ; The optimal individual Indi best The corresponding optimal SAR image recognition model is transmitted to the SAR image online recognition module;

[0090] (13) The SAR image data preprocessing module obtains the SAR image data in the real-time monitoring process from the SAR system, and obtains the online detection data set X after image data normalization and image data normalization operations. t , the online detection dataset X t Transmit to the SAR image online recognition module;

[0091] (14) The optimal SAR image recognition model is obtained by online deployment of the SAR image online recognition module. The normal sample recognition accuracy of the optimal SAR image recognition model is calculated using the online detection dataset according to formula (10). And based on the fast gradient sign attack (FGSM), gradient iteration based attack (PGD), optimization based attack (CW), decision boundary based attack (Deepfool), jitter based attack (Jitter) and random noise based attack (Random Noise) to generate adversarial sample sets, the recognition accuracy of the online deployment model for these six adversarial samples is calculated according to formula (11)

[0092] The beneficial effects of the present invention are as follows: compared with the existing technology, the present invention not only realizes the automated optimization design and automated generation of the optimal deep neural network model with compromised multiple performance indicators for the SAR image recognition system, but also constructs a multi-feature weighted loss function that expands the feature combination method of the SAR image recognition model. While improving the recognition accuracy and adversarial robustness of the model, it also greatly reduces the number of model parameters, better realizes the lightweight of the model, and is more convenient for the online deployment and engineering implementation of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a schematic diagram of the principle of a robust SAR image recognition system and method integrating feature loss and three-objective optimization;

[0094] Figure 2 It is a schematic diagram of the depth-wise separable convolution operation, where the convolution kernel size is n×n;

[0095] Figure 3 It is a schematic diagram of the dilated convolution operation, which shows the variations of the dilated convolution kernel when the convolution kernel size is 3×3, 5×5, and 7×7;

[0096] Figure 4 It is a schematic diagram of the spatially separable convolution operation with kernel sizes of 1×n and n×1;

[0097] Figure 5 It is a schematic diagram of the compression excitation network module;

[0098] Figure 6This is an example of a single individual encoding diagram. Its encoding information is: [5,0,6,0,2,0,0,0,10,2,8,1,2,4,2,2,7,3,0,5,2,6,6,5,2,0,1,1,2,1,3,2,9,2,7,0,6,1,4,3,7,3,5,5,11,3,10,5,1,0,1,1,0.2411,0.5823,0.0106,0.0441,0.4584,0.0085], where [5,0,6,0,2,0,0,0,10,2,8,1,2,4,2,2,7,3,0,5,2,6,6,5] is Normal Unit encoding, [2,0,1,1,2,1,3,2,9,2,7,0,6,1,4,3,7,3,5,5,11,3,10,5] is Reduction Unit encoding, [1,0,1,1] is Restype encoding, [0.2411,0.5823,0.0106,0.0741,0.4584] is Loss Weights encoding, and [0.0085] is lr encoding;

[0099] Figure 7 It is an example diagram of parent individual coding involved in the mixed differential mutation operation;

[0100] Figure 8 This is an example diagram of the specific implementation of the hybrid differential mutation operation during the offline optimization of the SAR image recognition model;

[0101] Figure 9 This is an example diagram of the hybrid differential crossover operation during the offline optimization process of the SAR image recognition model. The values ​​at the 3rd, 49th, and 57th positions are the values ​​of the corresponding positions in the parent individual, and the values ​​of the remaining positions are the values ​​of the corresponding positions in the mutant offspring individual.

[0102] Figure 10 It is the Pareto front obtained in the final evolutionary round after offline optimization of the SAR image recognition model;

[0103] Figure 11 This is a comparison of the recognition accuracy of the model generated by the present invention and the existing model on the FUSAR-Ship dataset for normal samples (Clean) and six types of adversarial attack samples, where ResNst50 and ResNst101 are manually designed neural network models;

[0104] Figure 12 Comparison of the number of parameters of the model automatically generated by the technology of the present invention and the model of the prior art; DETAILED DESCRIPTION

[0105] The present invention will be further described below with reference to the accompanying drawings, and the purpose and effects of the present invention will become more apparent.

[0106] The FUSAR-Ship image dataset is used as the implementation dataset for the robust SAR image recognition system and method based on fusion feature loss and three-objective optimization. Figure 1 A schematic diagram of a robust SAR image recognition system and method integrating feature loss and three-objective optimization is presented. A robust SAR image recognition system integrating feature loss and three-objective optimization is described, characterized in that the system comprises a SAR image data preprocessing module, an offline training and optimization module for an adversarial robust SAR image recognition model integrating multi-feature loss and three-objective optimization, and a SAR image online recognition module.

[0107] The SAR image data preprocessing module collects image data of the SAR system's real-time monitoring from the SAR system's real-time database. During this process, there is a threat of adversarial attacks from malicious adversaries, resulting in the presence of adversarial samples in the obtained real-time monitoring image data. The real-time monitoring image data is subjected to image data normalization and image data normalization processing to generate an online detection dataset, and the online detection dataset is transmitted to the SAR image online recognition module. The SAR image data preprocessing module collects historical image data of the SAR system from the SAR system's historical database, and after image data normalization and image data normalization processing, generates an offline dataset. The offline dataset is divided into an offline training set and an offline validation set in a ratio of 4:1, and then transmitted to an offline training and optimization module of an adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization.

[0108] The offline training and optimization module of the adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization first performs integer encoding on the neural network architecture parameters of the SAR image recognition model, and performs real number encoding on multiple loss weight parameters and training learning rates, which are collectively used as individual mixed encodings; then, an initialized candidate population is randomly generated, and the offline training set is trained based on the SAR image recognition model architecture corresponding to different individual encodings, the training learning rate, and the multi-feature fusion loss obtained by weighted calculation of multi-feature loss of cross entropy loss, similar shallow loss, similar deep loss, heterogeneous shallow loss, and heterogeneous deep loss; the trained SAR image recognition model is used to calculate the recognition accuracy of normal samples of the offline validation set and the number of model parameters, and the SAR image adversarial samples are generated through the PGD adversarial attack module. The adversarial sample recognition accuracy of the model is calculated; the normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters of the SAR image recognition model are used as fitness indicators of the individual three-objective fitness function evaluation submodule, and the fast non-dominated sorting, crowding distance calculation and fitness comprehensive evaluation submodules are used to sort the individuals in the initial candidate population, generate the parent population according to the population size, and select the optimal individual from the obtained Pareto front individuals; the parent population is input into the hybrid differential mutation module and the hybrid differential crossover module, and after the offspring population is generated, the parent population and the offspring population are merged to produce a new generation of candidate population; the above evolutionary process is repeated until the maximum number of evolutionary rounds is reached, thereby obtaining the final optimal individual; the optimal SAR image recognition model corresponding to the optimal individual is transmitted to the SAR image online recognition module;

[0109] The SAR image online recognition module deploys the optimal SAR image recognition model online, performs image recognition on the online detection dataset, and calculates the normal sample recognition accuracy. It then generates adversarial samples of the online detection dataset through the Fast Gradient Sign (FGSM) attack module, the Gradient Iteration (PGD) attack module, the Optimization (CW) attack module, the Decision Boundary (Deepfool) attack module, the Jitter (Jitter) attack module, and the Random Noise (Random Noise) attack module. Using the optimal SAR image recognition model deployed online, the model's classification accuracy for the above adversarial samples is calculated, thereby obtaining the model's adversarial sample recognition accuracy.

[0110] The robust SAR image recognition method integrating feature loss and three-objective optimization comprises the following steps:

[0111] (1) The SAR image data preprocessing module obtains the SAR image data stored in the historical monitoring process from the SAR system as the original data set, and marks it as O; the image data normalization module is used to perform image data normalization operation on O to obtain the normalized data set X; the image data normalization module is composed of a segmentation operation and a scaling operation; the segmentation operation divides the SAR image in the data set into regions. When the input SAR image contains multiple recognition targets, it is segmented into multiple SAR images so that each SAR image contains only a single recognition target. When the input SAR image contains only a single recognition target, no operation is performed; the scaling operation adjusts the size of the SAR image in the data set to keep it consistent with the internal parameters of the SAR image recognition model; the image data normalization operation is performed on the normalized data set X using formula (1) to obtain the offline data set X o ; Then, according to the ratio of 4:1, o Divide into offline training set X o_train and offline validation set X o_val ;

[0112]

[0113] (2) Setting the relevant parameter values ​​in the offline training and optimization module of the robust SAR image recognition model that integrates multi-feature loss and three-objective optimization, the parameter values ​​include the population size N = 20, the number of layers of the SAR image recognition model architecture combination Layers = 3, the number of units of the SAR image recognition model architecture combination layer c = 6, and the lower limit of the normal architecture unit parameter Normal low =[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0], the upper limit of the normal architecture unit parameter up =[11,1,11,1,11,2,11,2,11,3,11,3,11,4,11,4,11,5,11,5,11,6,11,6], constraining the lower limit of the architecture unit parameters low =[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0], constraining the upper limit of the architecture unit parameter reduction up =[11,1,11,1,11,2,11,2,11,3,11,3,11,4,11,4,11,5,11,5,11,6,11,6], the lower limit of the residual architecture control parameter Restype low =[0,0,0,0], the upper limit of the residual architecture control parameter Restype up =[1,1,1,1], the lower limit of the loss weight parameter Losslow =[0.01,0.01,0.01,0.01], the upper limit of the loss weight parameter Loss up = [1.0, 1.0, 1.0, 1.0], the lower limit of the learning rate parameter lr low = [0.0001], the upper limit of the learning rate parameter lr up =[0.01], training round EP train =20, mutation scaling factor σ = 0.8, crossover rate β = 0.9, maximum evolutionary round E max =20;

[0114] (3) Encode the architecture parameters as integers, and encode the loss weights and training learning rates as real numbers, which are used together as individual mixed encodings in the three-objective optimization method. Initialize 2N individuals as the initial candidate population Q h , where the encoding form of each individual is Indi = [Normal Unit, Reduction Unit, Restype, Loss Weights, lr], Indi represents any individual in the population, where Normal Unit, Reduction Unit and Restype together constitute the architecture parameters of the SAR image recognition model, Loss Weights represents the loss weight parameter of the SAR image recognition model, and lr represents the training learning rate of the SAR image recognition model;

[0115] The specific encoding forms of Normal Unit and Reduction Unit are the same, both are [(o1 1 ,n1 1 ,o2 1 ,n2 1 ),(o1 2 ,n1 2 ,o2 2 ,n2 2 ),...,(o1 c ,n1 c ,o2 c ,n2 c )],remember(o1 i ,n1 i ,o2 i ,n2 i ) is the i-th subunit, where 1≤i≤c; for the i-th subunit, n1 i Represents the first operation node in the subunit, o1 i Represents the operation node n1 i The specific operation, n2 i Represents the second operation node in the unit, o2 iRepresents the operation node n2 i Specific operation; j i is the specific operation on the jth node in the i-th unit, where 1≤j≤2, o j i The corresponding relationship with the specific operation is: when o j i = 0, perform the maximum pooling operation; when o j i =1, perform average pooling operation; when o j i = 2, no operation is performed; when o j i = 3, a depth-separable convolution operation with a convolution kernel size of 3×3 is performed; when o j i = 4, a depth-separable convolution operation with a convolution kernel size of 5×5 is performed; when o j i = 5, a dilated convolution operation with a kernel size of 3×3 is performed; when o j i = 6, a dilated convolution operation with a convolution kernel size of 5×5 is performed; when o j i =7, a dilated convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 8, a depth-wise separable convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 9, the spatially separable convolution operations with kernel sizes of 1×3 and 3×1 are performed; when o j i = 10, the spatially separable convolution operations with kernel sizes of 1×5 and 5×1 are performed; when o j i = 11, the spatially separable convolution operations with kernel sizes of 1×7 and 7×1 are performed respectively; for Normal Unit and Reduction Unit, the difference between the two is that when o j i When coming from Normal Unit, j i The parameters involved in the step length in the corresponding specific operations are all set to 1; when o j i When coming from Reduction Unit, o j i The parameters involving step length in the corresponding specific operations are all set to 2; Figure 2 A schematic diagram of the depth-wise separable convolution operation is given, Figure 3 A schematic diagram of the dilated convolution operation is given. Figure 4 A schematic diagram of the spatially separable convolution operation is given;

[0116] The correspondence between Normal Unit encoding and Reduction Unit encoding and model architecture is as follows: States is the set of intermediate feature values ​​of the model architecture corresponding to the current Unit encoding, and its initial value is States = [s1, s2], s1, s2 are the intermediate features of the model transferred to the current Unit; traverse the current Unit encoding in turn, according to n1 i and n2 i The value of selects the intermediate feature of the corresponding position in States, using o1 i and o2 i The corresponding operation processes the intermediate feature values ​​and merges the results to obtain the new model intermediate features and add them to States. Finally, the intermediate features that have not been selected in States except s1 and s2 are merged as the output of the model architecture corresponding to the current Unit encoding;

[0117] The specific form of Restype encoding is [Res1, Res2, ..., Res Layers ], used to control whether the model architecture corresponding to each Unit encoding adopts residual connection. k The model architecture corresponding to the k-th Unit encoding, where 1≤k≤Layers; when Res k = 0, the model architecture corresponding to the Unit encoding does not use residual connection, and the corresponding mode of the model architecture does not change; when Res k =1, the model architecture corresponding to the Unit encoding adopts residual connection, and the output of the Unit is merged with the initial feature s2 input to the Unit as the new output;

[0118] The correspondence between individual architecture codes and SAR image recognition model architectures is as follows: Based on the number of SAR image recognition model architecture combination layers (Layers) and the residual control code (Restype), the model architecture corresponding to the Normal Unit code and the model architecture corresponding to the Reduction Unit code are sequentially selected for concatenation. The model architecture corresponding to each Unit code is considered a layer of model architecture combination layer. After cross-stacking in the order of Normal Unit, Reduction Unit, Normal Unit, ReductionUnit, ..., Normal Unit, Reduction Unit, a neural network model with the number of layers is formed. A compression excitation network module is added after each layer of model architecture combination layer, and a global pooling layer is added after the last layer of model architecture combination layer. Figure 5 A schematic diagram of the compression excitation network module is given;

[0119] The specific form of Loss Weights encoding is [w1,w2,w3,w4,w5]; for the SAR image recognition model architecture corresponding to an individual, w1 is the cross entropy loss weight, w2 is the same shallow loss weight, w3 is the same deep loss weight, w4 is the different shallow loss weight, and w5 is the different deep loss weight, and their values ​​are all greater than 0;

[0120] The specific form of lr encoding is [lr]. For the SAR image recognition model architecture corresponding to an individual, lr is the learning rate of the model during offline training, and its value is greater than 0; Figure 6 An example diagram of a single individual encoding is given;

[0121] Initial candidate population Q h The specific generation process is as follows:

[0122] (3.1) According to the lower limit of the normal architecture unit parameter Normal low , the upper limit of the normal architecture unit parameter Normal up , constraining the lower limit of the architecture unit parameters Reduction low , constraining the upper limit of the architecture unit parameters Reduction up , the lower limit of the residual architecture control parameter Restype low , the upper limit of the residual architecture control parameter Restype up , the lower limit of the loss weight parameter Loss low , the upper limit of the loss weight parameter Loss up , the lower limit of the learning rate parameter lr low , the upper limit of the learning rate parameter lr up, generate N initial individuals Indis, the integer code value of the initial individual is a random integer between the lower and upper limits of the corresponding parameter, and the real number code value is a random real number between the lower and upper limits of the corresponding parameter;

[0123] (3.2) Note Indi up =[Normal up ,Reduction up ,Restype up ,Loss up ,lr up ], Indi low =[Normal low ,Reduction low ,Restype low ,Loss low ,lr low ]; the opposite individual of individual Indi in the coding space is recorded as Indi con ,Indi con The definition of is shown in formula (2); Generate N opposite individuals Indis con ;

[0124] Indi con =Indi low +Indi up -Indi (2)

[0125] (3.3) N initial individuals Indis and N opposite individuals Indis con A total of 2N individuals serve as the initial candidate population Q h ;

[0126] (4) Label the current evolutionary round of the multi-objective optimization technique as E, and set E = 0;

[0127] (5) For the candidate population Q h The offline training and optimization module of the adversarial robust SAR image recognition model integrating multi-feature loss and three-objective optimization is based on Q h The model architecture parameters corresponding to the individuals in the neural network model are used to generate the neural network model, and then the weighted loss of the model is calculated according to the Loss Weights parameter as the total loss of the model l sum , using the lr parameter as the training learning rate of the model, and using the adaptive moment estimation optimizer (Adam) to optimize the offline training set X o_train The model is trained with SAR image data in EP. train ;

[0128] The total loss of the model is l sumUse formula (3) to calculate, where l ce represents the cross entropy loss of the model, Represents the same shallow loss of the model, represents the same deep loss of the model, represents the heterogeneous shallow loss of the model, Represents the heterogeneous deep loss of the model. ce , Calculate according to formula (4) to formula (8) respectively;

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] Where θ represents all the parameters that constitute the current model, x represents the image data sample in the offline training set, and y represents the category corresponding to x. One-hot encoding is used. The implementation process of one-hot encoding is as follows;

[0136] (5.1) The correct category of the image is y true ,y n is the number of total classification categories;

[0137] (5.2) Generate a line of length y n The all-zero vector y one_hot ;

[0138] (5.3) one_hot Middle y true The value at the position is changed to 1;

[0139] (5.4)y one_hot That is category y true One-hot encoding of ;

[0140] Where n represents the total number of classification categories of samples in the offline training set, y (m) represents the mth value of y, f(x) (m) represents the m-th bit value in the prediction result of the current model for the image data sample x, that is, the prediction probability of the m-th class; p m Represents the number of image data samples of the mth category in the current batch training data, and Represents two different image data samples belonging to the mth category in the current batch training data, where f m ≠g m , and Represents the fth n Class and g n Two image data samples of the class, where f n ≠g n Where f surface (x) represents the shallow feature output of the current model for the image data sample x, that is, the feature value output by the first hidden layer of the current model, and the feature value is subjected to maximum pooling processing; f deep (x) represents the deep feature output of the current model for the image data sample x, that is, the eigenvalue output by the penultimate hidden layer of the current model, and the eigenvalue is subjected to maximum pooling processing; A(a,b) is the permutation number calculation, which represents the total number of permutation schemes for sequentially selecting b elements from a elements for permutation, where a and b are both positive integers, and the calculation process is shown in formula (9);

[0141]

[0142] After training, use the offline validation set X according to formula (10) o_val Calculate the normal sample recognition accuracy of the SAR image recognition model Based on the iterative gradient adversarial attack (PGD), the adversarial sample set of the offline verification set is generated, and the adversarial sample recognition accuracy of the model is calculated according to formula (11): where Num clean_right Represents the number of samples correctly identified in the offline validation set, Num adv_right Represents the number of correctly identified samples in the adversarial sample set of the offline validation set, Num all Represents the total number of samples in the offline validation set; counts the number of internal parameters of the SAR image recognition model Params;

[0143]

[0144]

[0145] (6) For the candidate population Q hThe individuals in the population are subjected to fast non-dominated sorting, crowding sorting and fitness comprehensive evaluation operations. The negative number of the individual's normal sample recognition accuracy is recorded as f1, the negative number of the individual's adversarial sample recognition accuracy is recorded as f2, and the number of model parameters Params is recorded as f3. The fast non-dominated sorting and crowding sorting select N individuals as the parent population Q. The fitness comprehensive evaluation operation calculates the fitness comprehensive evaluation value of the Pareto frontier individuals obtained after the fast non-dominated sorting, and selects the optimal individual. Specifically, the fast non-dominated sorting measures the three evaluation indicators of the individual's normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters, and divides the individuals to obtain the initial candidate population Q. h Pareto frontier individuals;

[0146] When individual Indi i The three optimization objective function values ​​of are all smaller than those of another individual Indi j When the individual is called Indi i Dominant individual Indi j . Remember n i is the number of individuals that dominate the i-th individual in the population, S i is the set of individuals dominated by the i-th individual; the specific implementation process of fast non-dominated sorting is:

[0147] (6.1) Let F1 be the empty set;

[0148] (6.2) Traverse the population Q in sequence h Individuals in, calculate the n corresponding to each individual i value, and will be controlled by the individual

[0149] Other individuals are placed in the corresponding set S i middle;

[0150] (6.3) n i = 0 individuals are put into the set F1;

[0151] (6.4) Let j1 = 1;

[0152] (6.5) Let H be the empty set;

[0153] (6.6) Traversal All individuals in , denoted by S u For The set of individuals dominated by the u-th individual, n q To dominate S u The number of individuals in the qth individual; get Each individual in S u ;

[0154] (6.7) For all S u , traverse each Su For all individuals in , calculate the n corresponding to each individual q value, and let n q =n q -1;

[0155] (6.8) If n q =0, then S u The qth individual is put into the set H;

[0156] (6.9) Let j1 = j1 + 1;

[0157] (6.10)

[0158] (6.11) Repeat steps (6.5) to (6.10) until you get is an empty set;

[0159] After completing the fast non-dominated sort, the parent population Q of size N is generated by combining the congestion sort. The specific implementation process is as follows:

[0160] (6.12) Let Q be the empty set, j2 = 1;

[0161] (6.13) Calculation The crowding distance of all individuals in . i for The crowding distance of the i-th individual in is calculated by using f1, f2 and f3 respectively. Sort the individuals in the , and take the three individuals with the largest f1 value, the largest f2 value and the largest f3 value as the boundary individuals. The crowding degree d i The crowding distance of all individuals except the boundary individuals is calculated according to formula (12). z i+1 express f of the i+1th individual z Target fitness value, f z i-1 express f of the i-1th individual z Target fitness value;

[0162]

[0163] (6.14) Let |Q| be the number of individuals in Q, for The number of individuals in . When All individuals in are put into Q. When Individuals are sorted from large to small according to the crowding distance. Select N-|Q| individuals in turn and put them into Q;

[0164] (6.15) Let j2 = j2 + 1;

[0165] (6.16) Repeat steps (6.12) to (6.15) until |Q| = N;

[0166] After that, the fitness comprehensive evaluation operation is performed to select the best individual. The individuals in F1 are taken as Pareto frontier individuals, and the best individual is selected by combining the three optimization objective function values. The specific selection process is as follows: when F1 contains only one individual, the individual is selected as the best individual in this evaluation. best When F1 contains two or more individuals, the comprehensive fitness evaluation value f of the i-th individual in F1 is calculated according to formula (13): i all , where f1 i represents the optimization target value of f1 for the i-th individual, f2 i represents the optimization target value of f2 for the i-th individual, f3 i represents the f3 optimization target value of the i-th individual. Select f i all The smallest individual is the optimal individual Indi best , when there are multiple individuals with the smallest f i all When selecting The individual with the higher value is regarded as the optimal individual of the model best ;

[0167]

[0168] (7) For each individual in the population Q, a hybrid differential mutation module is used to perform a mutation operation, and the mutated population is marked as Q m , the specific implementation process is:

[0169] (7.1) Let i1 = 0, Q m is an empty set, and the i1th individual in Q is Indi i1 ;

[0170] (7.2) To India i1 Randomly select three other individuals Indi r1 ,Indi r2 and Indi r3 , where i1≠r1≠r2≠r3;

[0171] (7.3) According to the fitness comprehensive evaluation operation in step (6), select Indi r1 ,Indi r2 and Indi r3 The best individual in rb ;

[0172] (7.4) i1 The mutant offspring individual is recorded as Indi muti1 Indi(Arch) represents the value of the integer code position in the individual Indi code, Indi(Other) represents the value of the real code position in the individual Indi code, Indi muti1 Integer encoding part Indi muti1 The variation value of (Arch) is calculated by formula (14), Indi muti1 (Other) The variation value of the real number coding part is calculated by formula (15), where σ represents the variation scaling factor, Represents a floor operation;

[0173]

[0174] Indi muti1 (Other)=Indi rb (Other)+σ×(Indi r2 (Other)-Indi r3 (Other)) (15)

[0175] (7.5)Indi muti1 (Arch) and Indi muti1 (Other) after splicing to get Indi muti1 , will Indi muti1 The values ​​exceeding the boundary value in the code are modified to the boundary value;

[0176] (7.6) Let i1 = i1 + 1;

[0177] (7.7) Repeat steps (7.2) to (7.6) until every individual in Q has undergone mutation. Put all the mutated offspring individuals into the population Q m middle; Figure 7 An example diagram of parent individual coding involved in mixed differential mutation operation is given. Figure 8 An example diagram of the specific implementation of the mixed differential mutation operation is given; the integer encoding part is calculated according to formula (14) and rounded down, and the real number encoding part is calculated according to formula (15);

[0178] (8) For population Q mFor each individual in the , a hybrid difference crossover module is used for crossover operation, and the population after crossover is marked as Q x , the specific implementation process is:

[0179] (8.1) Let i2 = 0, Q x is an empty set, let Q m The i2th individual is Indi muti2 , the i2th individual in Q is Indi i2 ;

[0180] (8.2) Note Indi muti2 The crossover offspring individual is Indi ci2 ,Indi muti2 The encoding length is l Indi2 . Generate a 1 to l Indi2 A uniformly distributed random integer in the range k1, where 1≤k1≤l Indi2 , let k1 be the fixed cross position;

[0181] (8.3) Note Indi ci2 (k2) is Indi ci2 The value of the k2th position on the , Rand(k2) generates a uniformly distributed random number in the range of 0 to 1, where 0≤Rand(k2)≤1, and Indi is calculated according to formula (16) ci2 The value of each bit in ;

[0182]

[0183] (8.4) The individual Indi ci2 The values ​​outside the boundary value in the code are modified to the boundary value;

[0184] (8.5) Let i2 = i2 + 1;

[0185] (8.7) Repeat steps (8.2) to (8.5) until Q m Each individual in the crossover operation is performed. All the obtained crossover offspring individuals are placed into the population Q x middle; Figure 9 An example of the concrete implementation of hybrid differential crossover operation is given;

[0186] (9) Combine population Q with population Q x Merge into a new candidate population Q hnew ;

[0187] (10) Let Q h =Q hnew , E=E+1;

[0188] (11) Repeat steps (5) to (10) until E = E max ;

[0189] (12) For the E max Candidate population Q for rounds of evolution h According to the steps (5) and (6), fast non-dominated sorting and fitness comprehensive evaluation are performed to obtain the optimal individual Indi best ; The optimal individual Indi best The corresponding optimal SAR image recognition model is transmitted to the SAR image online recognition module; Figure 10 The Pareto frontier of the final evolutionary round is given; the final selected Indi best The encoding is: [5,0,6,0,2,0,0,0,10,2,8,1,2,4,2,2,7,3,0,5,2,6,6,5,2,0,1,1,2,1,3,2,9,2,7,0,6,1,4,3,7,3,5,5,11,3,10,5,1,0,1,1,0.2411,0.5823,0.0106,0.0741,0.4584,0.0085];

[0190] (13) The SAR image data preprocessing module obtains the SAR image data in the real-time monitoring process from the SAR system, and obtains the online detection data set X after image data normalization and image data normalization operations. t , the online detection dataset X t Transmit to the SAR image online recognition module;

[0191] (14) The optimal SAR image recognition model is obtained by online deployment of the SAR image online recognition module. The normal sample recognition accuracy of the optimal SAR image recognition model is calculated using the online detection dataset according to formula (10). And based on the fast gradient sign attack (FGSM), gradient iteration based attack (PGD), optimization based attack (CW), decision boundary based attack (Deepfool), jitter based attack (Jitter) and random noise based attack (Random Noise) to generate adversarial sample sets, the recognition accuracy of the online deployment model for these six adversarial samples is calculated according to formula (11)

[0192] The SAR image recognition model automatically generated by the present invention is experimentally tested and compared with two existing technologies (ResNst50 and ResNst101) on the FUSAR-Ship dataset. The experimental test selected six types of adversarial samples involved in the present invention, including FGSM attack, PGD attack, CW attack, Deepfool attack, Jitter attack and Randrom Noise attack; the recognition accuracy of the present invention and the existing RseNet50 and RseNet101 technologies on normal samples (Clean) and six types of adversarial samples is compared. Figure 11 As shown in the figure, it is not difficult to find that: compared with the existing technology, the technology of the present invention has the highest recognition accuracy for normal samples and the recognition accuracy for the six types of adversarial attack samples. In addition, the number of model parameters of the SAR image recognition model automatically generated by the technology of the present invention is compared with the existing RseNet50 and RseNet101 technologies. Figure 12 As shown in the figure, it can be clearly seen that the number of parameters of the SAR image recognition model automatically generated by the technology of the present invention is also much lower than that of the existing RseNet50 and RseNet101 technologies.

[0193] In summary, the beneficial effects of the present invention are as follows: compared with the existing technology, the present invention not only realizes the automated optimization design and automated generation of the optimal deep neural network model with multiple performance indicators compromise for the SAR image recognition system, but also constructs a multi-feature weighted loss function that expands the feature combination method of the SAR image recognition model. While improving the recognition accuracy and adversarial robustness of the model, it also greatly reduces the number of model parameters, better realizes the lightweight of the model, and is more convenient for the online deployment and engineering implementation of the model.

Claims

1. A robust synthetic aperture radar image recognition system integrating feature loss and three-objective optimization, characterized in that: The system includes a SAR image data preprocessing module, an offline training and optimization module for an adversarial robust SAR image recognition model integrating multi-feature loss and three-objective optimization, and a SAR image online recognition module. The SAR image data preprocessing module collects image data of the SAR system real-time monitoring from the SAR system real-time database. During this process, there is a threat of adversarial attacks from malicious adversaries, resulting in the presence of adversarial samples in the obtained real-time monitoring image data. The real-time monitoring image data is processed by image data normalization and image data normalization to generate an online detection data set, and the online detection data set is transmitted to the SAR image online recognition module; The SAR image data preprocessing module collects historical image data of the SAR system from the SAR system historical database, generates an offline dataset after image data normalization and image data normalization processing, divides the offline dataset into an offline training set and an offline validation set in a ratio of 4:1, and then transmits it to the offline training and optimization module of the adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization; The offline training and optimization module of the adversarial robust SAR image recognition model that integrates multi-feature loss and three-objective optimization first performs integer encoding on the neural network architecture parameters of the SAR image recognition model, and performs real number encoding on multiple loss weight parameters and training learning rates, which are collectively used as individual mixed encodings; then, an initialized candidate population is randomly generated, and the offline training set is trained based on the SAR image recognition model architecture corresponding to different individual encodings, the training learning rate, and the multi-feature fusion loss obtained by weighted calculation of multi-feature loss of cross entropy loss, similar shallow loss, similar deep loss, heterogeneous shallow loss, and heterogeneous deep loss; the trained SAR image recognition model is used to calculate the recognition accuracy of normal samples of the offline validation set and the number of model parameters, and the SAR image adversarial samples are generated through the PGD adversarial attack module. The adversarial sample recognition accuracy of the model is calculated; the normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters of the SAR image recognition model are used as fitness indicators of the individual three-objective fitness function evaluation submodule, and the fast non-dominated sorting, crowding distance calculation and fitness comprehensive evaluation submodules are used to sort the individuals in the initial candidate population, generate the parent population according to the population size, and select the optimal individual from the obtained Pareto front individuals; the parent population is input into the hybrid differential mutation module and the hybrid differential crossover module, and after the offspring population is generated, the parent population and the offspring population are merged to produce a new generation of candidate population; the above evolutionary process is repeated until the maximum number of evolutionary rounds is reached, thereby obtaining the final optimal individual; the optimal SAR image recognition model corresponding to the optimal individual is transmitted to the SAR image online recognition module; The SAR image online recognition module deploys the optimal SAR image recognition model online, performs image recognition on the online detection dataset, and calculates the normal sample recognition accuracy. Then, adversarial samples of the online detection dataset are generated through the fast gradient sign (FGSM) attack module, the gradient iteration (PGD) attack module, the optimization (CW) attack module, the decision boundary (Deepfool) attack module, the jitter attack module, and the random noise (Random Noise) attack module. The optimal SAR image recognition model deployed online is used to calculate the model's classification accuracy for the above adversarial samples, thereby obtaining the model's adversarial sample recognition accuracy.

2. A robust SAR image recognition method using the fusion feature loss and three-objective optimization of the system of claim 1, characterized in that: The following steps are involved: (1) The SAR image data preprocessing module obtains the SAR image data stored in the historical monitoring process from the SAR system as the original data set and marks it as O; The image data normalization module is used to perform image data normalization operation on O to obtain a normalized dataset X; the image data normalization module consists of a segmentation operation and a scaling operation; the segmentation operation divides the SAR image in the dataset into regions. When the input SAR image contains multiple recognition targets, it is segmented into multiple SAR images so that each SAR image contains only a single recognition target. When the input SAR image contains only a single recognition target, no operation is performed; the scaling operation adjusts the size of the SAR image in the dataset to keep it consistent with the internal parameters of the SAR image recognition model; the image data normalization operation is performed on the normalized dataset X using formula (1) to obtain the offline dataset X o ; Then, according to the ratio of 4:1, o Divide into offline training set X o_train and offline validation set X o_val ; (2) Setting the relevant parameter values ​​in the offline training and optimization module of the robust SAR image recognition model that integrates multi-feature loss and three-objective optimization, the parameter values ​​include the population size N, the number of layers of the SAR image recognition model architecture combination Layers, the number of units in the SAR image recognition model architecture combination layer c, and the lower limit of the normal architecture unit parameter Normal low , the upper limit of the normal architecture unit parameter Normal up , constraining the lower limit of the architecture unit parameters Reduction low , constraining the upper limit of the architecture unit parameters Reduction up , the lower limit of the residual architecture control parameter Restype low , the upper limit of the residual architecture control parameter Restype up , the lower limit of the loss weight parameter Loss low , the upper limit of the loss weight parameter Loss up , the lower limit of the learning rate parameter lr low , the upper limit of the learning rate parameter lr up , training round EP train , mutation scaling factor σ, crossover rate β, maximum evolutionary round E max ; (3) Encode the architecture parameters as integers, and encode the loss weights and training learning rates as real numbers, which are used together as individual mixed encodings in the three-objective optimization method. Initialize 2N individuals as the initial candidate population Q h , where the encoding form of each individual is Indi = [Normal Unit, Reduction Unit, Restype, Loss Weights, lr], Indi represents any individual in the population, where Normal Unit, Reduction Unit and Restype together constitute the architecture parameters of the SAR image recognition model, Loss Weights represents the loss weight parameter of the SAR image recognition model, and lr represents the training learning rate of the SAR image recognition model; The specific encoding forms of Normal Unit and Reduction Unit are the same, both are [(o1 1 ,n1 1 ,o2 1 ,n2 1 ),(o1 2 ,n1 2 ,o2 2 ,n2 2 ),...,(o1 c ,n1 c ,o2 c ,n2 c )],remember(o1 i ,n1 i ,o2 i ,n2 i ) is the i-th subunit, where 1≤i≤c; For the i-th subunit, n1 i Represents the first operation node in the subunit, o1 i Represents the operation node n1 i The specific operation, n2 i Represents the second operation node in the unit, o2 i Represents the operation node n2 i Specific operation; j i is the specific operation on the jth node in the i-th unit, where 1≤j≤2, o j i The corresponding relationship with the specific operation is: when o j i = 0, perform the maximum pooling operation; when o j i =1, perform average pooling operation; when o j i = 2, no operation is performed; when o j i = 3, a depth-separable convolution operation with a convolution kernel size of 3×3 is performed; when o j i = 4, a depth-separable convolution operation with a convolution kernel size of 5×5 is performed; when o j i = 5, a dilated convolution operation with a kernel size of 3×3 is performed; when o j i = 6, a dilated convolution operation with a convolution kernel size of 5×5 is performed; when o j i =7, a dilated convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 8, a depth-wise separable convolution operation with a convolution kernel size of 7×7 is performed; when o j i = 9, the spatially separable convolution operations with kernel sizes of 1×3 and 3×1 are performed; when o j i = 10, the spatially separable convolution operations with kernel sizes of 1×5 and 5×1 are performed; when o j i = 11, the spatially separable convolution operations with kernel sizes of 1×7 and 7×1 are performed respectively; for Normal Unit and Reduction Unit, the difference between the two is that when o j i When coming from Normal Unit, j i The parameters involved in the step length in the corresponding specific operations are all set to 1; when o j i When coming from Reduction Unit, o j i The parameters involving step length in the corresponding specific operations are all set to 2; The correspondence between Normal Unit encoding and Reduction Unit encoding and model architecture is as follows: States is the set of intermediate feature values ​​of the model architecture corresponding to the current Unit encoding, and its initial value is States = [s1, s2], where s1 and s2 are the intermediate features of the model transferred to the current Unit; Traverse the current Unit code in sequence, according to n1 i and n2 i The value of selects the intermediate feature of the corresponding position in States, using o1 i and o2 i The corresponding operation processes the intermediate feature values ​​and merges the results to obtain a new model intermediate feature and adds the feature to States; finally, the intermediate features that have not been selected in States except s1 and s2 are merged as the output of the model architecture corresponding to the current Unit encoding; The specific form of Restype encoding is [Res1, Res2, ..., Res Layers ], used to control whether the model architecture corresponding to each Unit encoding adopts residual connection; k The model architecture corresponding to the k-th Unit encoding, where 1≤k≤Layers; when Res k = 0, the model architecture corresponding to the Unit encoding does not use residual connection, and the corresponding mode of the model architecture does not change; when Res k =1, the model architecture corresponding to the Unit encoding adopts residual connection, and the output of the Unit is merged with the initial feature s2 input to the Unit as the new output; The correspondence between individual architecture codes and SAR image recognition model architectures is as follows: according to the number of layers of the SAR image recognition model architecture combination layer (Layers) and the residual control code (Restype), the model architecture corresponding to the Normal Unit code and the model architecture corresponding to the Reduction Unit code are sequentially selected for splicing; the model architecture corresponding to each Unit code is regarded as a layer of model architecture combination layer, and after cross-stacking in the order of Normal Unit, Reduction Unit, Normal Unit, ReductionUnit, ..., Normal Unit, Reduction Unit, a neural network model with the number of layers is formed. A compression excitation network module is added after each layer of the model architecture combination layer, and a global pooling layer is added after the last layer of the model architecture combination layer; The specific form of Loss Weights encoding is [w1,w2,w3,w4,w5]; for the SAR image recognition model architecture corresponding to an individual, w1 is the cross entropy loss weight, w2 is the same shallow loss weight, w3 is the same deep loss weight, w4 is the different shallow loss weight, and w5 is the different deep loss weight, and their values ​​are all greater than 0; The specific form of lr encoding is [lr]; for the SAR image recognition model architecture corresponding to the individual, lr is the learning rate of the model during offline training, and its value is greater than 0; Initial candidate population Q h The specific generation process is as follows: (3.1) According to the lower limit of the normal architecture unit parameter Normal low , the upper limit of the normal architecture unit parameter Normal up , constraining the lower limit of the architecture unit parameters Reduction low , constraining the upper limit of the architecture unit parameters Reduction up , the lower limit of the residual architecture control parameter Restype low , the upper limit of the residual architecture control parameter Restype up , the lower limit of the loss weight parameter Loss low , the upper limit of the loss weight parameter Loss up , the lower limit of the learning rate parameter lr low , the upper limit of the learning rate parameter lr up , generate N initial individuals Indis, the integer code value of the initial individual is a random integer between the lower and upper limits of the corresponding parameter, and the real number code value is a random real number between the lower and upper limits of the corresponding parameter; (3.2) Note Indi up =[Normal up ,Reduction up ,Restype up ,Loss up ,lr up ], Indi low =[Normal low ,Reduction low ,Restype low ,Loss low ,lr low ]; the opposite individual of individual Indi in the coding space is recorded as Indi con ,Indi con The definition of is shown in formula (2); Generate N opposite individuals Indis con ; Now con =Now low +Now up -Now (2) (3.3) N initial individuals Indis and N opposite individuals Indis con A total of 2N individuals serve as the initial candidate population Q h ; (4) Label the current evolutionary round of the multi-objective optimization technique as E, and set E = 0; (5) For the candidate population Q h The fitness function is evaluated; the offline training and optimization module of the adversarial robust SAR image recognition model integrating multi-feature loss and three-objective optimization is based on Q h The model architecture parameters corresponding to the individuals in the neural network model are used to generate the neural network model, and then the weighted loss of the model is calculated according to the Loss Weights parameter as the total loss of the model l sum , using the lr parameter as the training learning rate of the model, and using the adaptive moment estimation optimizer (Adam) to optimize the offline training set X o_train The model is trained with SAR image data in EP. train ; The total loss of the model is l sum Use formula (3) to calculate, where l ce represents the cross entropy loss of the model, Represents the same shallow loss of the model, represents the same deep loss of the model, represents the heterogeneous shallow loss of the model, represents the heterogeneous deep loss of the model; l ce , Calculate according to formula (4) to formula (8) respectively; Where θ represents all the parameters that constitute the current model, x represents the image data sample in the offline training set, and y represents the category corresponding to x. One-hot encoding is used. The implementation process of one-hot encoding is as follows; (5.1) The correct category of the image is y true ,y n is the number of total classification categories; (5.2) Generate a line of length y n The all-zero vector y one_hot ; (5.3) one_hot Middle y true The value at the position is changed to 1; (5.4)y one_hot That is category y true One-hot encoding of ; Where n represents the total number of classification categories of samples in the offline training set, y (m) represents the mth value of y, f(x) (m) represents the m-th bit value in the prediction result of the current model for the image data sample x, that is, the prediction probability of the m-th class; p m Represents the number of image data samples of the mth category in the current batch training data, and Represents two different image data samples belonging to the mth category in the current batch training data, where f m ≠g m , and Represents the fth n Class and g n Two image data samples of the class, where f n ≠g n ; where f surface (x) represents the shallow feature output of the current model for the image data sample x, that is, the feature value output by the first hidden layer of the current model, and the feature value is subjected to maximum pooling processing; f deep (x) represents the deep feature output of the current model for the image data sample x, that is, the eigenvalue output by the penultimate hidden layer of the current model, and the eigenvalue is subjected to maximum pooling processing; A(a,b) is the permutation number calculation, which represents the total number of permutation schemes for sequentially selecting b elements from a elements for permutation, where a and b are both positive integers, and the calculation process is shown in formula (9); After training, use the offline validation set X according to formula (10) o_val Calculate the normal sample recognition accuracy of the SAR image recognition model Based on the iterative gradient adversarial attack (PGD), the adversarial sample set of the offline verification set is generated, and the adversarial sample recognition accuracy of the model is calculated according to formula (11): where Num clean_right Represents the number of samples correctly identified in the offline validation set, Num adv_right Represents the number of correctly identified samples in the adversarial sample set of the offline validation set, Num all Represents the total number of samples in the offline validation set; counts the number of internal parameters of the SAR image recognition model Params; (6) For the candidate population Q h The individuals in the population are subjected to fast non-dominated sorting, crowding sorting, and fitness comprehensive evaluation operations; the negative of the individual's recognition accuracy for normal samples is recorded as f1, the negative of the individual's recognition accuracy for adversarial samples is recorded as f2, and the number of model parameters Params is recorded as f3. The fast non-dominated sorting and crowding sorting select N individuals as the parent population Q. The fitness comprehensive evaluation operation calculates the fitness comprehensive evaluation value of the Pareto frontier individuals obtained after the fast non-dominated sorting, and selects the optimal individual; Specifically, fast non-dominated sorting measures the three evaluation indicators of individual normal sample recognition accuracy, adversarial sample recognition accuracy and number of model parameters, and divides the individuals to obtain the initial candidate population Q h Pareto frontier individuals; When individual Indi i The three optimization objective function values ​​of are all smaller than those of another individual Indi j When the individual is called Indi i Dominant individual Indi j ; Remember n i is the number of individuals that dominate the i-th individual in the population, S i is the set of individuals dominated by the i-th individual; the specific implementation process of fast non-dominated sorting is: (6.1) Let F1 be the empty set; (6.2) Traverse the population Q in sequence h Individuals in, calculate the n corresponding to each individual i value, and put the other individuals dominated by this individual into the corresponding set S i middle; (6.3) n i = 0 individuals are put into the set F1; (6.4) Let j1 = 1; (6.5) Let H be the empty set; (6.6) Traverse F j1 All individuals in , denoted by S u To be F j1 The set of individuals dominated by the u-th individual, n q To dominate S u The number of individuals in the qth individual; get F j1 Each individual in S u ; (6.7) For all S u , traverse each S u For all individuals in , calculate the n corresponding to each individual q value, and let n q =n q -1; (6.8) If n q =0, then S u The qth individual is put into the set H; (6.9) Let j1 = j1 + 1; (6.10) Let F j1 =H; (6.11) Repeat steps (6.5) to (6.10) until the obtained F j1 is an empty set; After completing the fast non-dominated sort, the parent population Q of size N is generated by combining the congestion sort. The specific implementation process is as follows: (6.12) Let Q be the empty set, j2 = 1; (6.13) Calculate F j2 The crowding distance of all individuals in i F j2 The crowding distance of the i-th individual in F is calculated by using f1, f2 and f3 respectively. j2 Sort the individuals in the , and take the three individuals with the largest f1 value, the largest f2 value and the largest f3 value as the boundary individuals. The crowding degree d i is considered infinite; except for the boundary individuals, the crowding distance of the remaining individuals is calculated according to formula (12); where f z i+1 Indicates F j2 f of the i+1th individual z Target fitness value, f z i-1 Indicates F j2 f of the i-1th individual z Target fitness value; (6.14) Let |Q| be the number of individuals in Q, for The number of individuals in When All individuals in are put into Q; when When Individuals are sorted from large to small according to the crowding distance. Select N-|Q| individuals in turn and put them into Q; (6.15) Let j2 = j2 + 1; (6.16) Repeat steps (6.12) to (6.15) until |Q| = N; After that, the fitness comprehensive evaluation operation is performed to select the best individual; the individuals in the obtained F1 are taken as the Pareto frontier individuals, and the best individual is selected by combining the three optimization objective function values. The specific selection process is: when F1 contains only one individual, the individual is selected as the best individual in this evaluation. best When F1 contains two or more individuals, the comprehensive fitness evaluation value f of the i-th individual in F1 is calculated according to formula (13): i all , where f1 i represents the optimization target value of f1 for the i-th individual, f2 i represents the optimization target value of f2 for the i-th individual, f3 i Represents the optimization target value of f3 of the i-th individual; select f i all The smallest individual is the optimal individual Indi best , when there are multiple individuals with the smallest f i all When selecting The individual with the higher value is regarded as the optimal individual of the model best ; (7) For each individual in the population Q, a hybrid differential mutation module is used to perform a mutation operation, and the mutated population is marked as Q m , the specific implementation process is: (7.1) Let i1 = 0, Q m is an empty set, and the i1th individual in Q is Indi i1 ; (7.2) To India i1 Randomly select three other individuals Indi r1 ,Indi r2 and Indi r3 , where i1≠r1≠r2≠r3; (7.3) According to the fitness comprehensive evaluation operation in step (6), select Indi r1 ,Indi r2 and Indi r3 The best individual in rb ; (7.4) i1 The mutant offspring individual is recorded as Indi muti1 ; Indi(Arch) represents the value of the integer code position in the individual Indi code, Indi(Other) represents the value of the real code position in the individual Indi code, Indi muti1 Integer encoding part Indi muti1 The variation value of (Arch) is calculated by formula (14), Indi muti1 (Other) The variation value of the real number coding part is calculated by formula (15), where σ represents the variation scaling factor, Represents a floor operation; Indi muti1 (Other)=Indi rb (Other)+σ×(Indi r2 (Other)-Indi r3 (Other)) (15) (7.5)Indi muti1 (Arch) and Indi muti1 (Other) after splicing to get Indi muti1 , will Indi muti1 The values ​​exceeding the boundary value in the code are modified to the boundary value; (7.6) Let i1 = i1 + 1; (7.7) Repeat steps (7.2) to (7.6) until every individual in Q has undergone mutation. Put all the mutated offspring individuals into the population Q. m middle; (8) For population Q m For each individual in the , a hybrid difference crossover module is used for crossover operation, and the population after crossover is marked as Q x , the specific implementation process is: (8.1) Let i2 = 0, Q x is an empty set, let Q m The i2th individual is Indi muti2 , the i2th individual in Q is Indi i2 ; (8.2) Note Indi muti2 The crossover offspring individual is Indi ci2 ,Indi muti2 The encoding length is l Indi2 ; Generate a 1 to l Indi2 A uniformly distributed random integer in the range k1, where 1≤k1≤l Indi2 , let k1 be the fixed cross position; (8.3) Note Indi ci2 (k2) is Indi ci2 The value of the k2th position on the , Rand(k2) generates a uniformly distributed random number in the range of 0 to 1, where 0≤Rand(k2)≤1, and Indi is calculated according to formula (16) ci2 The value of each bit in ; (8.4) The individual Indi ci2 The values ​​outside the boundary value in the code are modified to the boundary value; (8.5) Let i2 = i2 + 1; (8.7) Repeat steps (8.2) to (8.5) until Q m Each individual in the crossover operation is performed; all the obtained crossover offspring individuals are placed in the population Q x middle; (9) Combine population Q with population Q x Merge into a new candidate population Q hnew ; (10) Let Q h =Q hnew , E=E+1; (11) Repeat steps (5) to (10) until E = E max ; (12) For the E max Candidate population Q for rounds of evolution h According to the steps (5) and (6), fast non-dominated sorting and fitness comprehensive evaluation are performed to obtain the optimal individual Indi best ; The optimal individual Indi best The corresponding optimal SAR image recognition model is transmitted to the SAR image online recognition module; (13) The SAR image data preprocessing module obtains the SAR image data in the real-time monitoring process from the SAR system, and obtains the online detection data set X after image data normalization and image data normalization operations. t , the online detection dataset X t Transmit to the SAR image online recognition module; (14) The optimal SAR image recognition model is obtained by online deployment of the SAR image online recognition module. The normal sample recognition accuracy of the optimal SAR image recognition model is calculated using the online detection dataset according to formula (10). And based on the fast gradient sign attack (FGSM), gradient iteration based attack (PGD), optimization based attack (CW), decision boundary based attack (Deepfool), jitter based attack (Jitter) and random noise based attack (Random Noise) to generate adversarial sample sets, the recognition accuracy of the online deployment model for these six adversarial samples is calculated according to formula (11)

Images