CN114859316A

CN114859316A - Radar target intelligent identification method based on task relevance weighting

Info

Publication number: CN114859316A
Application number: CN202210669794.5A
Authority: CN
Inventors: 简涛; 刘传辉; 卢仁伟; 黄晓冬; 王海鹏; 但波; 谢梓铿
Original assignee: Naval Aeronautical University
Current assignee: Naval Aeronautical University
Priority date: 2022-06-14
Filing date: 2022-06-14
Publication date: 2022-08-05

Abstract

The invention discloses a radar target intelligent identification method based on task relevance weighting, and belongs to the field of radar signal processing. Respectively designing and identifying model loss functions in a pre-training stage of a meta-training task and a fine-tuning stage of a meta-testing task in a targeted manner; the fuzzy truncated cosine element training loss function in the pre-training stage emphasizes the generalization performance of the model and focuses on the general characteristics of extracting source domain target data; in the fine adjustment stage, the multi-class balance cosine included angle constraint loss function emphasizes balancing the occupation ratio of different classes in the loss function, and focuses on extracting the individual characteristics of the small sample target data to be identified; the model migration efficiency and the target domain identification accuracy are improved; the designed task loss negative exponential function weighting updating mode of the meta-learner parameters reduces the instability of the parameter updating process, and the calculating process is simple and convenient; the method improves the accuracy of small sample identification and has popularization and application values.

Description

Radar target intelligent identification method based on task relevance weighting

One, the technical field

The invention belongs to the field of radar signal processing, and particularly relates to a radar target intelligent identification method based on task relevancy weighting.

Second, background art

The High Resolution Range Profile (HRRP) is a projection of a target main scattering point in a radar irradiation direction, reflects relative position information of the target scattering point, reflects partial structure information of the target to a certain extent, and has the advantages of easy storage, easy processing, convenient acquisition and the like, so that the HRRP has a wide application prospect in the field of target identification. For cooperative targets, a large number of target HRRP samples can be acquired, but in practical applications, the target targeted by the radar is often a non-cooperative target, and it is difficult to acquire a sufficient number of non-cooperative target HRRP data. Therefore, the identification of the radar target HRRP under the condition of a small sample has become one of the research hotspots in the field of radar target identification.

The target classification and identification technology under the condition of an early small sample mainly takes a HRRP statistical modeling method as a main part, such as a linear dynamic model, a factor analysis model, a self-adaptive Gaussian classifier model and the like, then students utilize a multi-task learning concept and category label information to improve on the basis of the models, and a multi-task factor analysis model, a label auxiliary factor analysis model, a multi-task label constraint convolution factor analysis model and the like are successively provided. The statistical modeling method needs to make prior assumptions on the variables of the model, and when the actual environment is mismatched with the prior assumption conditions, the identification performance of the method is sharply reduced.

In recent years, the deep learning method has gained more and more attention in the field of radar target identification. The deep learning model can autonomously learn the intrinsic rules and expression levels of data, but the deep learning model often depends on a large amount of sample data and a deeper network structure to obtain the expected recognition effect. In an actual confrontation environment, the amount of acquired non-cooperative target data is small, the angle is often single, and for the small sample data, model overfitting can be caused by deep learning of an excessively deep network structure. Aiming at the problem, on one hand, a correlative scholars perform data enhancement by methods such as generation of a countermeasure network and the like from the viewpoint of increasing the data volume so as to meet the requirements of a deep learning model on a data set; on the other hand, a learning method based on model migration is researched, model parameters learned by a source domain are migrated to a target domain, and therefore small sample identification is carried out by using migrated knowledge. It should be noted that, in the model migration method, when the source domain and the target domain are different greatly, the created migration model is difficult to adapt to different recognition problems. As an extension of the migration learning method, meta-learning is a mechanism for designing learners with different characteristics aiming at different tasks, can guide the learning of a new task by using the experience on different tasks in the past, and is suitable for various application occasions such as classification, regression, reinforcement learning and the like.

It should be noted that the existing meta-learning method usually refers to the learning experiences of all tasks indiscriminately, neglects the difference influence of different task experiences on new tasks, and is easy to cause excessive reference to the low-relevancy experience, thereby causing the decrease of the recognition rate. In addition, because the source domain data and the target domain data are often greatly different, the pre-training model needs to have good generalization performance and can effectively extract the general characteristics of the source domain target data; for the meta-learning method combining the basic learner and the meta-learner, the loss function of the basic learner is easy to converge to a local minimum value, and the common cross entropy loss function can cause the model to obtain an excessively self-confident classification result, so that the basic learner model in the pre-training stage lacks necessary generalization performance. In addition, in the design of the existing meta-learning method combining the basic learner and the meta-learner, in order to simplify the difficulty of model design, the same loss function is often adopted in the meta-training task recognition training model and the meta-testing task recognition training model, while the potential difference between the source domain data corresponding to the meta-training task and the target domain data corresponding to the meta-testing task is ignored, the differential loss function design cannot be performed on different recognition tasks of the meta-training task and the meta-testing task, and the model migration efficiency and the target domain recognition accuracy are reduced. In fact, in the pre-training stage corresponding to the meta-training task, the design of the loss function in the recognition model should pay attention to the generalization performance of the model and pay attention to the general characteristics of the extracted source domain target data; under the condition of small samples, the situation that the number of samples is not uniform often exists among a plurality of target categories to be identified, in the process of back propagation, the proportion of categories with more samples in a loss function is high, the model is easy to be guided to be optimized in the direction which is favorable for outputting the classification result of the categories, and then the identification rate of the categories with less samples is reduced.

Aiming at the problem of poor identification precision of multiple classes of radar targets HRRP under the condition of small samples, how to design a targeted model in a pre-training stage and a fine-tuning stage is an important problem. In the pre-training stage of the meta-training task, how to design a basic learner loss function with strong generalization capability, and further carrying out targeted migration reference on learning experiences of different tasks in the meta-learner, introducing the difference learning experiences of different meta-training tasks in a smoother mode, guiding a new recognition task to carry out effective small sample recognition, and providing common experience for subsequent target recognition; in the fine-tuning stage of the meta-test task, how to design an identification model loss function capable of effectively extracting the individual characteristics of the small sample target data; the two stages are mutually cooperated, the task to be recognized is guided to carry out effective small sample recognition by means of meta-training task experience, the recognition precision of the target to be recognized is improved, and the method is one of the difficulties in the field of intelligent recognition of the multi-class small samples of the radar target.

Third, the invention

1. Technical problem to be solved

Aiming at the problem of poor multi-class identification precision of the radar target HRRP under the condition of a small sample, how to construct a proper meta-learning identification framework, decouple the characteristics of a single task and the commonality among the tasks, design a proper basic learner and a proper meta-learner to carry out experience migration on the characteristics of the single task and the commonality among the tasks, and respectively carry out targeted identification model loss function design in a pre-training stage and a fine-tuning stage so as to realize effective small sample identification of the radar target; aiming at the construction of a basic learner, how to design a loss function with strong generalization capability, and avoiding the defects that a common cross entropy loss function is easy to cause a model to obtain an excessively self-confident classification result and is easy to converge to a local minimum value, so that the general characteristics of source domain target data are effectively extracted, and necessary strong experience to be migrated with strong universality is reserved for a subsequent meta-test task; aiming at the construction of a meta-learner, how to design a smoother parameter updating mode is to excavate the difference characteristics of the learning experiences of different tasks, so that the targeted smooth migration reference of the learning experiences of different tasks is realized, the reference is used for guiding a new recognition task to carry out effective small sample recognition, and the small sample target recognition precision of the new task is improved; aiming at the construction of the identification model of the meta-test task, how to design a proper loss function, the problem of reduced identification rate of the class target with fewer samples caused by unbalanced number of different classes of samples is solved, the individual characteristics of the small sample target data are effectively extracted, and the matching degree of the meta-test task identification model to the task to be identified is enhanced.

2. Technical scheme

The invention discloses a radar target intelligent identification method based on task relevance weighting, which comprises the following steps:

step 1, constructing a data set of deep meta-learning; constructing a meta-test task small sample training data set and a meta-test task test data set according to the radar high-resolution range profile data of the N types of targets to be recognized, and classifying the tasks to be recognized into the meta-test task set; according to M types of non-to-be-recognized target accumulated data which are different from the to-be-recognized target data or different in data source, constructing a meta-training task set by randomly extracting N types of target data, and taking all the N types of non-to-be-recognized target data as a meta-training task training data set;

the meta-learning method specifically includes a neural network adaptation method, a metric learning adaptation method, a basic learner and meta-learner adaptation method, a bayesian meta-learning adaptation method, and the like. The basic learner and the meta learner are combined to decouple task characteristic modeling and task commonality modeling, and the model can achieve the optimal task characteristic and commonality through mutual communication between the basic learner and the meta learner, so that the generalization capability of the model is improved on the basis of keeping model precision. The data structure of the traditional machine learning is generally divided into a training set and a test set, and both the training set and the test set need to contain a large amount of sample data. The deep meta-learning is similar to the traditional machine learning and is also divided into a meta-training set and a meta-testing set, and the difference is that the meta-training set and the meta-testing set are not sample data but are sets of tasks, and each task comprises a corresponding training data set and a corresponding testing data set. Different from traditional machine learning and classical meta learning, the data structure of the meta learning model is divided into a meta training task set and a meta testing task set, the meta testing task set comprises a corresponding training data set and a corresponding testing data set, and in order to improve the data utilization rate, the meta training task set only comprises a corresponding training data set.

For the identification tasks of N types of radar targets, constructing a small sample training data set and a meta test task test data set of the meta test task according to radar high-resolution range profile data of the N types of targets to be identified, wherein the meta test task training data set is a small sample data set, and the tasks to be identified are classified into the meta test task set; according to M types of non-to-be-recognized target accumulated data of different target types or different data sources from the to-be-recognized target data, wherein M is larger than or equal to N, a meta-training task set is constructed by randomly extracting N types of target data from the M types, each meta-training task comprises N types of non-to-be-recognized targets, and each meta-training task only comprises a meta-training task training data set.

Different from the classical meta-learning method which divides meta-training task data into a training data set and a test data set, the invention takes all data in the meta-training task as the meta-training task training data set, improves the training utilization rate of the meta-training task data, improves the calculation precision and stability of characteristic parameters and loss values of a single meta-training task, is beneficial to providing more reliable task learning experience for a subsequent meta-learner, and is used for guiding the accurate design of a target small sample classifier to be recognized.

It is emphasized that, in view of the small sample characteristic of the target data to be recognized, the difference of the accumulated data of the non-target to be recognized of the present invention compared with the radar high-resolution range profile data of the N types of targets to be recognized can be embodied in two aspects: or different target types (for example, the target to be identified is a ship target, and accumulated data of the targets not to be identified relates to an airplane target); or the data types are different (for example, the data are also ship targets, the radar high-resolution range profile data of the target to be identified is measured data, and the accumulated data of the target not to be identified is simulation data); or the two aspects are different (for example, the radar high-resolution range profile data of the target to be identified is actually measured data, and the non-target to be identified is a ground vehicle target, and the accumulated data is one-dimensional radial sampling data of an infrared image or radar high-resolution range profile simulation data).

Step 2, designing a meta-learning model; designing a meta-learning model as a combination of a fuzzy truncation cosine loss basic learner and a task loss negative exponent correction meta-learner; the fuzzy truncated cosine loss basic learning device adopts a fuzzy truncated cosine element training loss function as a convolutional neural network loss function of an element training task; the task loss negative index correction meta-learner corrects the parameter update of the meta-learner by adopting the negative index function value of the meta-training task loss value based on the smooth average absolute error loss function; the input of the meta-learning model is a meta-training task set, and the output of the meta-learning model is updated model parameters;

the meta-learning model of the invention is designed as a combination of a fuzzy truncated cosine loss basis learner and a task loss negative exponent correction meta-learner. The basic learner models the task characteristics, requiring the basic learner to discover the inherent regularity of each task data set. The deep learning model depends on strong expression capability, can extract the deep features of the target, and has good fitting degree to data. The convolutional neural network is a typical representation of a deep learning model, and automatically extracts radar target data features by utilizing convolutional layers and constrains the features through a loss function. When the quantity of the target data to be recognized is sufficient, the model can be directly trained from the beginning, and the separability of the target features is further improved through a loss function. When the number of HRRP samples of the target to be recognized is lack, the model under-fitting phenomenon can be caused by the de novo training, and in order to solve the problem, the model can be pre-trained by using other target actual measurement data sets or simulation data sets which are irrelevant to the target data, namely, the model is pre-trained by using a meta-training task, and then the pre-trained model is finely adjusted by using the target data to be recognized corresponding to the training data sets in the meta-testing task. Therefore, the pre-training model needs to have better generalization performance, namely the model has the capability of extracting general features of the target data.

Aiming at the radar target HRRP multi-class small sample identification task, the invention designs a fuzzy truncated cosine loss basic learning device, takes a one-dimensional Convolutional Neural Network (CNN) as a learning training model, and constructs a fuzzy truncated cosine element training loss function as a loss function of the one-dimensional CNN. The fuzzy truncated cosine loss basis learner comprises an input layer, three convolutional layers (Conv), three pooling layers (POOL), a batch normalization layer (BN), two full-link layers (FC) and an output layer. The concrete connection mode is as follows:

Conv1→POOL1→Conv2→POOL2→Conv3+BN→POOL3→FC1→FC2

wherein, the input layer imports a small batch of data in each meta-training task training data set; the first two convolutional layers, Conv1 and Conv2, are standard convolutional layers; the third convolutional layer Conv3 is a standardized convolutional layer, a batch normalization layer BN is added on the basis of the convolutional layer Conv3 and is used for enhancing the gradient change of a fuzzy truncated cosine element training loss function, and the three convolutional layers all adopt one-dimensional convolutional kernels with the sizes of A and A respectively ₁ 、A ₂ And A ₃ The number of the same is B ₁ 、B ₂ And B ₃ Step length is D, and the filling modes are the same; the pooling layer POOL is an average pooling layer,the step length is H; the number of the neurons of the first full connection layer FC1 is F, and the number of the neurons of the second full connection layer FC2 is equal to the number N of target categories to be identified; the output layer calculates the class probability of each sample by utilizing a softmax function, and calculates a loss function value according to the calculation result.

Note that the radar target HRRP has attitude angle sensitivity, and the HRRP waveforms of different attitude angles of the same target are greatly different; the HRRP samples corresponding to part of attitude angles contain more scattering point information and are easy to identify, while the HRRP samples corresponding to the other part of attitude angles contain less scattering point information and are difficult to identify, but the HRRP samples of all the attitude angles are equally important for target identification, and the generalization performance of the model is determined. In order to ensure that the pre-training model can well extract the HRRP complete attitude angle invariant feature, the output probability of the class corresponding to the HRRP sample with high recognition difficulty needs to be improved, and the classical cross entropy loss function is difficult to meet the requirement. Aiming at the problems, the invention designs a fuzzy truncated cosine element training loss function as a loss function of a one-dimensional CNN in a basic learner. The fuzzy truncated cosine element training loss function contains two main parts. The first part is the fuzzy cosine loss function L _F The method is used for solving the problem of excessive confidence of the classification result of the model, and the output result is fuzzified to reduce the difference of the output of each neuron, so that the output result of each neuron plays a role in the transmission process, and the phenomenon of excessive confidence of the model is avoided. The second part is a truncated cosine loss function L _T The method is used for solving the problem that the output probability of the HRRP corresponding to the category with high identification difficulty is low, only the output result meeting the condition is subjected to back propagation by utilizing the truncation characteristic of the truncation function, and the weight of the output result of the sample data of the part is increased through equivalence, so that the model can better extract the characteristics of the target which is easy to be confused.

The fuzzy truncated cosine element training loss function is equivalent to a linear combination of the fuzzy cosine loss function and the truncated cosine loss function. The fuzzy truncated cosine loss function for each meta-training task can be expressed as:

wherein the fuzzy cosine loss function L _F Is composed of

Truncating the cosine loss function L _T Is composed of

In the three formulas, m is the number of sample data in the gradient updating process; x is the number of _i Representing a full-connection layer output characteristic vector corresponding to the ith sample; y is _i Is the true tag of the ith sample;

representing the output result of the ith sample in the output layer; w _j The j-th column of the weight matrix W of the full connection layer represents a weight vector corresponding to the j-th class target; log (log) ₂ (. cndot.) represents a base-2 logarithmic function; a. beta, s and alpha are positive real parameters; wherein the positive parameter a satisfies 0 ≦ a ≦ 1-cos (2 π/N), which is for enhancing the feature vector x _i And y _i Weight vector W corresponding to class target _yi The included angle of (2) is restrained; the real parameter s solves the problem of non-convergence of the loss function and satisfies

ε is a positive constant close to 0; the positive parameter beta is more than or equal to 0 and less than or equal to 1 and is used for adjusting the output weight of the fuzzy loss component; the real parameter alpha is used for adjusting the weight of the truncated cosine loss function; phi is a _(·,·) Representing the included angle between two vectors with the same dimension;

representing a truncation function.

Using a truncation function

The reverse propagation of the output result only meeting the condition can be realized, and the truncated cosine loss function L is properly added _T The weight alpha of the model can better extract the characteristics of the confusable target; wherein the truncation function

Can be expressed as:

wherein T represents a truncation threshold, and T is more than or equal to 0.5 and less than or equal to 1; u (-) represents a modified step function, which can be expressed as

Where x represents the argument of the modified step function. Note L _t Only if the output result satisfies

And the model can better extract the characteristics of the confusable target by equivalently increasing the weight of the output result of corresponding sample data.

The fuzzy truncated cosine element training loss function is based on large edge cosine loss, and can continuously reduce the angle distance of the features in the class and improve the separability of the features while constraining the features between the classes. If the angle between the two vectors x and y is phi _(x,y) If the cosine distance between the two vectors x and y is the cosine value of the angle between the two vectors x and y, the inner product of the normalized vectors can be expressed as cos (phi) _(x,y) )＝x ^T y/(| x | | | y |), superscript T represents the transposition, | · | | represents to take the Euclidean norm, and the cosine distance can effectively avoid the adverse effect of the magnitude of the modulus of different vectors on the distance measurement.

Is a weight vector W _j And a feature vector x _i The similarity between the two is measured because the cosine value of the included angle is between

The interval is monotonously decreased, the included angle between the vectors is smaller, the rest chord value is larger, and the feature vector x is _i The greater the probability that the corresponding ith sample belongs to the category j; in order to further increase the inter-class distance of different classes of target features, the invention introduces a hyper-parameter a when the requirement is met

When the feature vector x is enhanced, the ith sample is judged to belong to the category j, and the hyperparameter a has the function of enhancing the feature vector x _i And weight vector

Is constrained.

The meta-learner in the meta-learning model design mainly models the commonality among tasks, and has the function of summarizing and concluding the training experience on all tasks after the training of the basic learner is finished each time, synthesizing new experience and feeding back the new experience to the basic learner. The meta learner has various forms, can be any model based on random gradient descent, and can also be a parameter updating algorithm. In order to reduce the operation complexity and improve the parameter updating precision and efficiency, the method designs a task loss negative exponent correction element learner, based on a smooth average absolute error loss function, the negative exponent function value of the element training task loss value is adopted to correct the element learner parameter updating, the task loss factor is introduced through the smooth negative exponent function, the difference influence of different task experiences on a new task is fully utilized, the learning experiences of different element training tasks are pertinently borrowed, the problem of identification rate reduction caused by over-borrowing of low-correlation experiences is avoided, and the initialization model parameter theta of the element learner is enabled to be reduced _M And updating the approach to the common direction of the tasks with high relevance.

The main purpose of meta-learning is to find the common direction of the high relevance task, so that the initial model parameter theta of the meta-learner is _M Approaching characteristic parameters of high-relevance element training tasks in the basic learner, namely obtaining a parameter theta _M The error distance between the characteristic parameters of all the meta-training tasks is minimized, i.e. the learning objective function of the meta-learner is as follows

Wherein D is _k Training a characteristic parameter theta of a task for a kth element _Bk And the meta learner parameter θ _M The error between. Different from the design of an objective function of a common meta-learner, inequality measures are omitted in the formula (6), only main error items are reserved, on the basis of ensuring the parameter updating precision, the subsequent gradient solving difficulty and the calculation complexity of the meta-learner parameter updating are reduced, and the parameter updating efficiency is improved.

In order to prevent the problem of gradient explosion of mean square error (corresponding to Euclidean space distance) and avoid the defect that the gradient of the mean absolute error at the zero point is not smooth, the method constructs a smooth mean absolute error loss function. Error term D _k Is particularly shown as

The learning goal of the meta-learner is to minimize the error between the two learner parameters. From the Lagrange theorem, D can be calculated _k At the parameter theta _M To minimize D _k . For theta _Bk -θ _M ||<1, for the distance D _k The derivation is carried out to obtain:

meta-learner often assumes everyIndividual element training task pair element learning device parameter theta _M The influence degrees of the parameters are the same, and then all the tasks are matched with the meta-learner parameter theta _M The effects of (a) are superimposed indifferently for computing the direction of commonality of all tasks. Considering the difference influence of different element training task experiences on the new task, the task loss value L of the kth element training task _k The difference degree of learning experience can be reflected to a certain extent, the smaller the loss value is, the larger the negative index function value is, the better the training effect is, namely, the learning experience of the task is more worthy of reference; and conversely, the larger the loss value is, the smaller the negative exponential function value is, the poorer the learning effect of the element training task is, and the lower the learning experience can be used for reference. Therefore, in order to avoid over-reference to the low-relevancy experience and fully utilize the deviation influence of different meta-training tasks on the parameter updating of the meta-learner, the method provided by the invention utilizes the negative exponential function value of the loss value of the meta-training task to perform weighted correction on the parameter updating process of the meta-learner, so that the meta-learner can pertinently reference the experience of each meta-training task, the over-reference to the low-relevancy experience is reduced, and the reference degree of the high-relevancy experience is improved. The parameter updating mode of the task loss negative exponent correcting element learner is as follows:

wherein,

i.e. updated parameter of meta learner, epsilon ₀ Is the update step length of the meta learner; due to a negative exponential function e ^-x The change along with the independent variable is relatively smooth, so that the adverse effect on the parameter updating process caused by the severe change of the task loss values of different training tasks can be effectively reduced; and the value range is (0, 1)]Effectively avoid L _k The problem of divergence of the parameter updating process caused by undersize or oversize is solved, and the calculation process is simple, convenient and fast and is convenient to realize.

Step 3, randomly extracting K meta-training tasks from the meta-training task set to carry out basic learningTraining a machine; for the kth element training task, carrying out batch training on a fuzzy truncated cosine loss basis learning device by using an element training task training data set, and obtaining a characteristic parameter theta of the kth element training task through iterative updating optimization of a convolutional neural network _Bk And loss value L _k (ii) a Forward transmitting the characteristic parameters and loss values of the K element training tasks to a task loss negative exponent correction element learner, and turning to the step 4;

randomly extracting K meta-training tasks from the meta-training task set to perform basic learner training; for the kth element training task, the fuzzy truncated cosine loss basic learning device designed in the step 2 is utilized, the training data set of the kth element training task is combined to carry out batch training, iterative updating optimization training is carried out through a convolutional neural network according to the fuzzy truncated cosine element training loss function of the formula (1), and the characteristic parameter theta of the kth element training task is obtained _Bk And corresponding task loss value L _k (ii) a After all the K element training tasks finish the training of the basic learner, the characteristic parameters { theta ] of the K element training tasks are used _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK And corresponding task loss value L _B1 ,L _B2 ,…,L _Bk ,…,L _BK And (5) transmitting the positive direction to the task loss negative index correction element learner, and turning to the step 4 to update the element learner parameters.

Step 4, the task loss negative index correction meta-learner uses the negative index function value of the corresponding meta-training task loss value to correct the meta-learner parameter theta based on the characteristic parameters of the K meta-training tasks _M Performing weighted correction update, and updating the updated parameters

Feeding back to a fuzzy truncated cosine loss basis learner; to be provided with

Performing a new round of optimization training for the new initialization model parameters of the fuzzy truncation cosine loss basic learner in the step 3 again;

task loss negative exponent correction element learningCharacteristic parameter theta of training task based on K elements _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK Using corresponding meta-training task loss value { L } _B1 ,L _B2 ,…,L _Bk ,…,L _BK The negative exponential function value of (9) is based on the element learning parameter theta _M Performing weighted correction update, and updating the updated parameters

step 5 executes step 3 and step 4 in a loop until N is reached _c The secondary loop is finished, and the last task loss negative exponent correction element learner parameter is saved

Constructing a convolutional neural network identification model with the same structure as a fuzzy truncated cosine loss basic learning device, and adopting a multi-class equilibrium cosine included angle constraint loss function as a convolutional neural network loss function of a unit test task; will be provided with

As an initialization parameter of the convolutional neural network identification model, introducing a small sample training data set of the element test task in batches to perform training and updating of the identification model, and finishing the training of the identification model when the multi-class equilibrium cosine included angle constraint loss function of the convolutional neural network is converged and does not decrease any more; and classifying and identifying the meta-test task test data set by using the trained convolutional neural network identification model, and evaluating the classification and identification accuracy of the N types of targets to be identified.

In the fine adjustment stage corresponding to the meta-test task, the target identification accuracy under the condition of a small sample can be greatly improved by fine adjustment on the basis of the pre-training model with high generalization performance. Under the condition of small samples, the number of the samples is often unbalanced among a plurality of categories, in the process of back propagation, the categories with more samples account for higher proportion of a loss function, and the model is easily guided to be optimized in the direction which is favorable for outputting the category classification result, so that the identification rate of the category targets with less samples is reduced. Therefore, a loss function of the convolutional neural network identification model at the fine tuning stage corresponding to the meta-test task needs to be designed in a targeted manner.

The large edge cosine loss restrains the features between classes, so that the angular distance of the features in the classes is continuously reduced, and the separability of the features is improved. Therefore, a convolutional neural network identification model with the same structure as a fuzzy truncated cosine loss basic learning device is constructed to serve as a fine tuning model corresponding to the element test task, and a multi-class equilibrium cosine included angle constraint loss function is adopted to serve as a convolutional neural network loss function of the element test task; the method comprises the following steps that a multi-class equilibrium cosine included angle constraint loss function calculates the Focal loss on the basis of cosine distance, and the angle distance of features in a class is continuously reduced while features between the classes are constrained by utilizing the enhanced edge cosine loss, so that the separability of the features is improved; meanwhile, the loss function reduces the weight of the easily distinguished classes in the loss function by introducing the Focal loss, so that the occupation ratio of each class in the loss function is balanced, and the problem of reduced recognition rate of the class target with less sample number caused by unbalanced sample numbers of different classes is solved.

In the convolutional neural network recognition model of the fine tuning stage corresponding to the meta-test task, the multi-class equilibrium cosine included angle constraint loss function can be expressed as:

wherein, the parameter gamma is used for adjusting the weight of the output, and the gamma is more than or equal to 0 and less than or equal to 5.

The meta-learning training stage is structurally divided into an inner loop and an outer loop, wherein the inner loop refers to the cyclic updating of parameters between the basic learner and the meta-learner, and the outer loop refers to the training round loop of the inner loop. Firstly, initializing model parameters of a basic learner, extracting an identification task from a meta-training set, importing the identification task into the basic learner for learning, and updating the parameters of the basic learner according to a fuzzy truncated cosine meta-training loss function; and secondly, importing the model parameters learned by the basic learner into a task loss negative exponent correction meta-learner for parameter updating, feeding the updated parameters back to the basic learner for parameter assignment again, performing a new round of meta-training task extraction and basic learner training, repeating the steps until the outer circulation times are reached, and storing the final CNN model parameters.

In the meta-learning test stage, the training parameters in the meta-learning training stage are used for carrying out parameter initialization on the recognition model so as to transfer the past learning experience to a meta-test set, and the initialized recognition model is used for carrying out classification recognition to verify the efficiency of the method. Firstly, using CNN model parameters stored in a meta-learning training stage as initial parameters of an identification model, importing a small sample training data set in a meta-test set to perform identification model training until a multi-class equilibrium cosine included angle constraint loss function is converged, and finishing meta-test set training; and then, carrying out classification, identification and evaluation on the test data set of the meta-test set by using the identification model and parameters obtained by training the meta-test set, and calculating the classification, identification and accuracy of each type of target in the test data set.

3. Advantageous effects

Compared with the background art, the invention has the beneficial effects that:

1) aiming at the problem of poor multi-class identification precision of the radar target HRRP under the condition of small samples, a meta-learning identification framework of a fuzzy truncated cosine loss basic learner combined with a task loss negative index correction meta-learner is constructed, the characteristics of a single task and the commonality among the tasks are decoupled, the characteristics of the single task and the commonality among the tasks are subjected to experience migration, and the effective identification of the small samples of the radar target HRRP is realized.

2) The potential difference of source domain data corresponding to the meta-training task and target domain data corresponding to the meta-testing task is fully considered, and targeted identification model loss function design is respectively carried out in a pre-training stage of the meta-training task and a fine-tuning stage of the meta-testing task; in the pre-training stage, the generalization performance of the model is emphasized in the loss function design of the recognition model, and the universal characteristic of the target data of the source domain is emphasized; in the design of the loss function of the identification model in the fine adjustment stage, attention is paid to balance the proportion of different classes in the loss function, and the attention is paid to extracting the individual characteristics of the target data of the small sample to be identified; the model migration efficiency and the target domain identification accuracy are improved.

3) Aiming at the CNN construction of a basic learner, a fuzzy truncated cosine element training loss function is designed, on one hand, the problem of excessive confidence of a model classification result is solved by constructing the fuzzy cosine loss function, and the output result is fuzzified to reduce the output difference of each neuron, so that the output result of each neuron plays a role in the transmission process, and the phenomenon of excessive confidence of the model is avoided; on the other hand, a truncated cosine loss function is constructed to solve the problem that the output probability of the class corresponding to part of sample data with high identification difficulty is low, only the output result meeting the conditions is subjected to back propagation by utilizing the truncation characteristic of the truncated function, and the weight of the output result of the part of sample data is increased equivalently, so that the model can better extract the characteristics of the target easy to be confused; the two aspects have synergistic effect, the generalization capability of the fuzzy truncated cosine element training loss function is improved, the general characteristics of the source domain target data are effectively extracted, and necessary and strong experience to be migrated with universality is reserved for a subsequent element testing task.

4) A meta-learner learning objective function only containing main error items is constructed, and by neglecting unequal measurement with low influence degree, on the basis of ensuring the parameter updating precision, the subsequent gradient solving difficulty and the calculation complexity of meta-learner parameter updating are reduced, and the parameter updating efficiency is improved.

5) The method has the advantages that a smooth average absolute error loss function of the meta-learner is constructed, the problem of gradient explosion of mean square error is solved, the defect that the gradient of the average absolute error at a zero point is not smooth is overcome, the modeling precision of an error item of a target function of the meta-learner is improved, and the model guarantee is provided for efficient updating of the parameters of the subsequent meta-learner.

6) The task loss negative index function weighting updating mode of the meta-learner parameters is designed, the unstable influence of the parameter updating process caused by the violent change of the loss values of different meta-training tasks is reduced by utilizing the smoothness of the change of the negative index function, the problem of divergence of the parameter updating process caused by the undersize or oversize loss values of the meta-training tasks is avoided, the calculating process is simple, convenient and fast, and convenient to realize.

7) Aiming at a fine tuning CNN model of a meta-test task, a multi-class equilibrium cosine included angle constraint loss function is designed, and the angle distance of the features in the classes is continuously reduced while the features between the classes are constrained by utilizing the enhanced edge cosine loss, so that the separability of the features is improved; and simultaneously, the weight of the easy-to-distinguish classes in the loss function is reduced by introducing the Focal loss, so that the occupation ratio of each class in the loss function is balanced, the problem of reduced recognition rate of class targets with fewer samples caused by unbalanced samples of different classes is solved, and the matching degree of the meta-test task recognition model to the task to be recognized is enhanced.

Description of the drawings

FIG. 1 is a diagram of a meta-learning data structure of the method of the present invention;

FIG. 2 is a diagram of a meta-learning model structure of the method of the present invention;

FIG. 3 is a schematic diagram of the basic learner structure of the method of the present invention;

FIG. 4 is a flow chart of a radar target intelligent identification method based on task relevance weighting.

Fifth, detailed description of the invention

The invention is further described below with reference to the accompanying drawings. The present embodiments are to be considered as illustrative and not restrictive, and all changes and modifications that come within the spirit of the invention and the scope of the appended claims are desired to be protected.

In order to verify the effectiveness of the method, the specific implementation mode provides 2 embodiments, the first embodiment is applied to the HRRP identification of multiple ship target radars, and the second embodiment is applied to the HRRP identification of multiple airplane target radars.

Example 1:

the embodiment 1 is application of HRRP identification of 5-class ship target radars, and specifically comprises the following steps:

step A-1: constructing a deep meta-learning data set, wherein the meta-learning data structure of the deep meta-learning data set is shown in FIG. 1; firstly, acquiring simulation data of an HRRP (high resolution ratio) of an M-type 10 ship target, constructing a meta-training task set by randomly extracting N-type 5 target data, and taking all N-type non-to-be-identified ship target simulation data as a meta-training task training data set; in view of the non-cooperative property of the actual test ship target, only HRRP (high resolution ratio) measured data of small samples of the ship target to be identified can be obtained, a meta-test task small sample training data set and a meta-test task test data set are constructed according to the HRRP measured data of the radar of the N classes of ship targets to be identified, and the task to be identified is classified into the meta-test task set.

Step A-2: designing a meta learning model; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss basis learner and a task loss negative exponent correcting meta-learner, and the model structure is shown in fig. 2. The fuzzy truncated cosine loss basic learning device adopts a fuzzy truncated cosine element training loss function as a convolutional neural network loss function of an element training task; the task loss negative index correction meta-learner corrects the parameter update of the meta-learner by adopting the negative index function value of the meta-training task loss value based on the smooth average absolute error loss function; the input of the meta-learning model is a meta-training task set, and the output of the meta-learning model is updated model parameters;

step A-3: extracting 5 types of target data from simulation data of 10 types of ship targets HRRP without repetition to form a set consisting of 252 meta-training tasks, and randomly extracting K to 10 meta-training tasks from the meta-training task set to train a basic learner;

CNN model parameter theta for fuzzy truncation cosine loss basic learner _B Make initialization, concrete packageWeight of convolutional layer { w _conv1 ,b _conv1 ,w _conv2 ,b _conv2 ，w _conv3 ,b _conv3 }, weight of full connection layer { w _fc ,b _fc ,w _fc2 ,b _fc2 }. The structural parameter value of the fuzzy truncated cosine loss basic learning device is as follows: n is 5, A ₁ ＝9，A ₂ ＝9，A ₃ ＝9，B ₁ ＝8，B ₂ ＝16，B ₃ 32, D ═ 1, H ═ 2, F ═ 100; the specific design of the corresponding fuzzy truncated cosine loss basis learner is shown in fig. 3.

For the kth element training task, a fuzzy truncated cosine loss basic learning device is utilized, batch training is carried out by combining a training data set of the kth element training task, iterative optimization updating of parameters is carried out by utilizing an Adam optimizer according to a convolutional neural network fuzzy truncated cosine element training loss function of the formula (1), and then a characteristic parameter theta of the kth element training task is obtained _Bk And corresponding task loss value L _k (ii) a After all the K element training tasks finish the training of the basic learner, the characteristic parameters { theta ] of the K element training tasks are used _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK And corresponding task loss value L _B1 ,L _B2 ,…,L _Bk ,…,L _BK And (5) transmitting the positive direction to the task loss negative index correction element learner, and turning to the step A-4 to update the element learner parameters.

Step A-4: task loss negative exponent correction element learning device trains characteristic parameters { theta ] of task based on K elements _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK Using corresponding meta-training task loss value { L } _B1 ,L _B2 ,…,L _Bk ,…,L _BK The negative exponential function value of (9) is based on the element learning parameter theta _M Performing weighted correction update, and updating the updated parameters

Feeding back to the fuzzy truncated cosine loss basic learner in the step A-3; to be provided with

Re-executing the step A-3 to perform a new round of optimization training for the new initialization model parameters of the fuzzy truncation cosine loss basic learner;

step A-5: the flow of the radar target intelligent identification method based on task relevance weighting is shown in fig. 4. According to the processing flow, circularly executing the step A-3 and the step A-4 until N is reached _c When 25 cycles are finished, the last meta learner parameter is saved

Constructing CNN recognition models with the same structure according to a fuzzy truncated cosine loss basic learning device

As an initialization parameter of the CNN recognition model, introducing a small sample training data set of a meta-test task in batches to perform model training and updating, and finishing model training when the CNN multi-class equilibrium cosine included angle constraint loss function represented by the formula (10) is converged and does not fall any more; and classifying and identifying the meta-test task test data set by using the trained CNN identification model, and evaluating the classification and identification accuracy of the measured data of the 5 types of ship targets to be identified.

Experimental analysis results show that compared with a CNN model, a deep migration learning model (DTL) and a meta learning model (UBML) with a task experience being used for unbiased reference, the meta learning model based on task relevancy weighting designed by the invention greatly improves the identification performance of multiple types of ship targets actually measured HRRP under small sample conditions by means of different targeted migration utilization of previous different simulated ship target identification task experiences. Under the condition of small samples, the CNN model directly carries out training test from beginning to end, and due to the reduction of training data, the CNN network is easy to generate overfitting and poor in identification rate; DTL can only migrate single task characteristics, the identification is limited by the experience quality of a single source task, and the identification effect of a small sample is poor; UBML does not refer to the learning experience of meta-training tasks differentially, and can perform more accurate classification and identification through the migration of task common experience, but the same CNN loss function is adopted in the pre-training stage and the fine-tuning stage, so that the extraction of the source domain target data general characteristic in the pre-training stage and the extraction of the individual characteristic of the small sample target data to be identified in the fine-tuning stage are difficult to realize, the identification rate of the small sample target of the ship is poor due to the excessive reference of the low-correlation experience, and the reduction of the identification rate of the class target with less sample number caused by the unbalanced sample number of different classes of targets is difficult to avoid; the meta-learning model based on task relevance weighting designed by the invention carries out task loss negative exponential function weighting correction on different task learning experiences in the meta-learning device, the method comprises the steps of introducing a loss value of a meta-training task in a smoother mode, guiding a new recognition task to carry out effective small sample recognition, utilizing a fuzzy truncated cosine meta-training loss function of a basic learner CNN model to extract general characteristics of source domain target data, utilizing a meta-testing task to finely tune a multi-class balanced cosine included angle constraint loss function of the CNN model, extracting individual characteristics of small sample target data to be recognized, balancing the occupation ratio of each class in the loss function, avoiding the problem of reduced recognition rate of class targets with fewer samples due to unbalanced number of different classes of samples, enhancing the matching degree of the meta-testing task recognition model to the recognition task, and improving the comprehensive migration efficiency of the model and the recognition accuracy of the target domain. In conclusion, the method is superior to other existing learning models in whole, is more suitable for classification and identification of small sample ship targets than the conventional transfer learning method, and embodies the superiority of the method.

Example 2:

embodiment 2 is an application of HRRP identification of a 4-class aircraft target radar, which may specifically include the following steps:

step B-1: a deep meta-learning data set is constructed, and the meta-learning data structure of the deep meta-learning data set is shown in FIG. 1. Due to the lack of simulation models of the airplane target, the accumulated ground vehicle target actual measurement HRRP data can be used for constructing a meta-learning training task. Firstly, acquiring measured data of an M-8 vehicle target HRRP, constructing a meta-training task set by randomly extracting N-4 target data, and taking all the N types of non-to-be-identified vehicle target measured data as a meta-training task training data set; in view of the non-cooperative property of the actual test airplane target, only HRRP (high resolution ratio) measured data of a small sample of the airplane target to be identified can be obtained, a meta-test task small sample training data set and a meta-test task test data set are constructed according to the HRRP measured data of the radar of the N types of airplane targets to be identified, and the task to be identified is classified into the meta-test task set.

Step B-2: designing a meta learning model; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss basis learner and a task loss negative exponent correcting meta-learner, and the model structure is shown in fig. 2. The fuzzy truncated cosine loss basic learning device adopts a fuzzy truncated cosine element training loss function as a convolutional neural network loss function of an element training task; the task loss negative index correction meta-learner corrects the parameter update of the meta-learner by adopting the negative index function value of the meta-training task loss value based on the smooth average absolute error loss function; the input of the meta-learning model is a meta-training task set, and the output of the meta-learning model is updated model parameters;

step B-3: extracting 4 types of target data from the actual measurement data of 8 types of vehicle targets HRRP without repetition to form a set consisting of 70 meta-training tasks, and randomly extracting K to 10 meta-training tasks from the meta-training task set to train a basic learner;

CNN model parameter theta for fuzzy truncation cosine loss basic learner _B Initializing, including the weight { w of convolutional layer _conv1 ,b _conv1 ,w _conv2 ,b _conv2 ，w _conv3 ,b _conv3 }, weight of full connection layer { w _fc ,b _fc ,w _fc2 ,b _fc2 }. The structural parameter value of the fuzzy truncated cosine loss basic learning device is as follows: n is 5, A ₁ ＝9，A ₂ ＝9，A ₃ ＝9，B ₁ ＝8，B ₂ ＝16，B ₃ 32, D ═ 1, H ═ 2, F ═ 100; the specific design of the corresponding fuzzy truncated cosine loss basis learner is shown in fig. 3.

For the k-th element training task, a fuzzy truncated cosine loss basic learning device is utilized, the training data set of the k-th element training task is combined for batch training, and the convolutional neural network fuzzy truncated cosine element training loss function according to the formula (1)And performing iterative optimization updating on the parameters by using an Adam optimizer to further obtain a characteristic parameter theta of the kth element training task _Bk And corresponding task loss value L _k (ii) a After all the K element training tasks finish the training of the basic learner, the characteristic parameters { theta ] of the K element training tasks are used _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK And corresponding task loss value L _B1 ,L _B2 ,…,L _Bk ,…,L _BK And (5) transmitting the positive direction to the task loss negative index correction element learner, and turning to the step B-4 to update the element learner parameters.

Step B-4: task loss negative exponent correction element learning device trains characteristic parameters { theta ] of task based on K elements _B1 ,θ _B2 ,…,θ _Bk ,…,θ _BK Using corresponding meta-training task loss value { L } _B1 ,L _B2 ,…,L _Bk ,…,L _BK The negative exponential function value of (9) is based on the element learning parameter theta _M Performing weighted correction update, and updating the updated parameters

Feeding back to the fuzzy truncated cosine loss basic learner in the step B-3; to be provided with

B-3 is executed again to carry out a new round of optimization training for the new initialization model parameters of the fuzzy truncation cosine loss basic learner;

step B-5: the flow of the radar target intelligent identification method based on task relevance weighting is shown in fig. 4. According to the processing flow, circularly executing the step B-3 and the step B-4 until N is reached _c End of 7 cycles, save last meta learner parameter

AsThe method comprises the steps of (1) identifying initialization parameters of a model, introducing a small sample training data set of a meta-test task in batches to perform training and updating of the model, and finishing model training when a CNN multi-class equilibrium cosine included angle constraint loss function represented by the formula (10) is converged and does not decrease any more; and classifying and identifying the meta-test task test data set by using the trained CNN identification model, and evaluating the classification and identification accuracy of the 4 types of airplane target actual measurement data to be identified.

Experimental analysis results show that compared with a CNN model, a deep migration learning model (DTL) and a meta learning model (UBML) for which task experience is not used for bias reference, the meta learning model based on task relevancy weighting greatly improves the identification performance of multiple types of airplane target actual measurement HRRPs under the condition of small samples by means of different targeted migration utilization of previous different actual measurement vehicle target identification task experiences. Under the condition of small samples, the CNN model directly carries out training test from beginning to end, and due to the reduction of training data, the CNN network is easy to generate overfitting and poor in identification rate; DTL can only migrate single task characteristics, the identification is limited by the experience quality of a single source task, and the identification effect of a small sample is poor; UBML does indifference reference to the learning experience of meta-training tasks, and can carry out more accurate classification and identification through the migration of task common experience, but the same CNN loss function is adopted in the pre-training stage and the fine-tuning stage, so that the extraction of the source domain target data general characteristics in the pre-training stage and the extraction of the individual characteristics of the small sample target data to be identified in the fine-tuning stage are difficult to realize, the identification rate of the small sample target of the airplane target is poor due to the excessive reference to the low-correlation experience, and the reduction of the identification rate of the target of the category with less sample number caused by the unbalanced sample number of different categories of targets is difficult to avoid; the meta-learning model based on task relevance weighting designed by the invention carries out task loss negative exponential function weighting correction on different task learning experiences in the meta-learning device, the method comprises the steps of introducing a loss value of a meta-training task in a smoother mode, guiding a new recognition task to carry out effective small sample recognition, utilizing a fuzzy truncated cosine meta-training loss function of a basic learner CNN model to extract general characteristics of source domain target data, utilizing a meta-testing task to finely tune a multi-class balanced cosine included angle constraint loss function of the CNN model, extracting individual characteristics of small sample target data to be recognized, balancing the occupation ratio of each class in the loss function, avoiding the problem of reduced recognition rate of class targets with fewer samples due to unbalanced number of different classes of samples, enhancing the matching degree of the meta-testing task recognition model to the recognition task, and improving the comprehensive migration efficiency of the model and the recognition accuracy of the target domain. In conclusion, the method is superior to other existing learning models in whole, is more suitable for classification and identification of small-sample airplane targets than the conventional transfer learning method, and embodies the superiority of the method.

Claims

1. The radar target intelligent identification method based on task relevance weighting is characterized by comprising the following steps:

step 1, constructing a data set of deep meta-learning; constructing a small sample training data set and a meta-test task test data set of a meta-test task according to radar high-resolution range profile data of N types of targets to be identified, and classifying the tasks to be identified into a meta-test task set; according to M types of non-to-be-recognized target accumulated data which are different from the to-be-recognized target data or different in data source, constructing a meta-training task set by randomly extracting N types of target data, and taking all the N types of non-to-be-recognized target data as a meta-training task training data set;

step 2, designing a meta-learning model; designing a meta-learning model as a combination of a fuzzy truncation cosine loss basic learner and a task loss negative exponent correction meta-learner; the fuzzy truncated cosine loss basic learning device adopts a fuzzy truncated cosine element training loss function as a convolutional neural network loss function of an element training task; the task loss negative index correction meta-learner performs weighted correction updating on meta-learner parameters by adopting negative index function values of meta-training task loss values based on a smooth average absolute error loss function; the input of the meta-learning model is a meta-training task set, and the output of the meta-learning model is updated model parameters;

step 3, randomly extracting K meta-training tasks from the meta-training task set to carry out basic learner training; for the k-th element training task, after fuzzy truncationOn the basis of the string loss learner, the training data set of the element training task is used for batch training, and the characteristic parameter theta of the kth element training task is obtained through iterative updating optimization of the convolutional neural network _Bk And loss value L _k (ii) a Forward transmitting the characteristic parameters and loss values of the K element training tasks to a task loss negative exponent correction element learner, and turning to the step 4;

As an initialization parameter of the convolutional neural network identification model, introducing a small sample training data set of the element test task in batches to perform training and updating of the identification model, and finishing the training of the identification model when the multi-class equilibrium cosine included angle constraint loss function of the convolutional neural network is converged and does not decrease any more; identification using trained convolutional neural networksAnd the model is used for carrying out classification and identification on the meta-test task test data set and evaluating the classification and identification accuracy of the N types of targets to be identified.

2. The radar target intelligent identification method based on task relevance weighting according to claim 1, wherein in the step 2:

in the construction of the convolutional neural network of the fuzzy truncated cosine loss basic learning device, a fuzzy truncated cosine element training loss function is designed and is formed by linear combination of the fuzzy truncated cosine loss function and the truncated cosine loss function, and the fuzzy truncated cosine element training loss function of each element training task is expressed as

Wherein m is the number of sample data in the gradient updating process; x is the number of _i Representing a full-connection layer output characteristic vector corresponding to the ith sample; y is _i Is the authentic label for the ith sample;

representing the output result of the ith sample in the output layer; w _j Is the weight vector corresponding to the jth class target; a. beta, s and alpha are positive real parameters; phi is a _(·,·) Representing the included angle between two vectors with the same dimension;

represents a truncation function, which can be expressed as

。

3. The radar target intelligent identification method based on task relevance weighting according to claim 1, wherein in the step 2:

the main error item in the learning objective function of the task loss negative exponential correction element learner is designed into a smooth average absolute error loss function, and the characteristic parameter theta of the kth element training task _Bk And the meta learner parameter θ _M Error D between _k Is shown as

Wherein, | | · | | represents taking euclidean norm.

4. The radar target intelligent identification method based on task relevance weighting according to claim 1, wherein in the step 2:

the parameter updating mode of the task loss negative exponent correcting element learner is

Wherein,

namely the updated parameters of the meta learner.

5. The radar target intelligent identification method based on task relevance weighting according to claim 1, characterized in that in the step 5:

the multi-class equilibrium cosine included angle constraint loss function balances the occupation ratio of each class in the loss function by reducing the weight of the easy-to-distinguish class in the loss function, and can be expressed as: