Intelligent target recognition method based on lossy fuzzy correction

Abstract

The present invention discloses a target intelligent recognition method based on loss fuzzy correction, which belongs to the field of radar signal processing. A recognition framework of a fuzzy truncated cosine loss base learner combined with a task loss inverse weighted update meta-learner is constructed. In the pre-training stage of the meta-training task, a fuzzy truncated cosine meta-training loss function is designed for the convolutional neural network construction of the base learner; in the fine-tuning stage of the meta-test task, a multi-class balanced cosine angle constraint loss function is designed for the fine-tuning convolutional neural network model; a task loss inverse weighted update method for the meta-learner parameters is designed, and the negative correlation between the meta-training task loss value and the task training effect is used to avoid excessive reference to low-correlation experience. The method of the present invention improves the model migration efficiency and the accuracy of multi-category recognition of high-resolution range images of radar targets under small sample conditions, and has popularization and application value.

Description

1. Technical Field

[0001] The invention belongs to the field of radar signal processing, and in particular relates to a target intelligent recognition method based on loss fuzzy correction. 2. Background Technology

[0002] High Resolution Range Profile (HRRP) is the projection of the main scattering point of the target in the direction of radar illumination. It reflects the relative position information of the target scattering point and, to a certain extent, reflects the partial structural information of the target. It has the advantages of easy storage, easy processing and convenient acquisition, so it has broad application prospects in the field of target recognition. For cooperative targets, a large number of target HRRP samples can be obtained. However, in practical applications, the targets targeted by radar are often non-cooperative targets, and it is difficult to obtain sufficient non-cooperative target HRRP data. Therefore, radar target HRRP recognition under small sample conditions has become one of the research hotspots in the field of radar target recognition.

[0003] In the early days, target classification and recognition technology under small sample conditions was mainly based on the statistical modeling method of HRRP, such as linear dynamic model, factor analysis model and adaptive Gaussian classifier model. Later, scholars used the concept of multi-task learning and category label information to improve on the above models, and successively proposed multi-task factor analysis model, label-assisted factor analysis model and multi-task label-constrained convolution factor analysis model. Statistical modeling methods require prior assumptions on the variables of the model. When the actual environment does not match the prior assumptions, the recognition performance of this type of method will drop sharply.

[0004] In recent years, deep learning methods have received more and more attention in the field of radar target recognition. Deep learning models can autonomously learn the inherent laws and representation levels of data, but they often need to rely on a large amount of sample data and a deeper network structure to obtain the expected recognition effect. In actual adversarial environments, the amount of non-cooperative target data that can be obtained is small and the angle is often relatively single. For the above small sample data, a deep network structure that is too deep in deep learning will lead to model overfitting. To address this problem, relevant scholars, on the one hand, from the perspective of increasing the amount of data, use methods such as generative adversarial networks to enhance data in order to meet the requirements of deep learning models for data sets; on the other hand, they study learning methods based on model transfer, which transfers the model parameters learned in the source domain to the target domain, so as to use the transferred knowledge for small sample recognition. It should be pointed out that in the model transfer method, when the source domain and the target domain are very different, the established transfer model is difficult to apply to different recognition problems. As an extension of the transfer learning method, meta-learning is a mechanism for designing learners with different characteristics for different tasks. It can use the experience of previous different tasks to guide the learning of new tasks. It is suitable for classification, regression, reinforcement learning and other application scenarios.

[0005] It should be pointed out that the existing meta-learning methods often indiscriminately borrow the learning experience of all tasks, ignoring the differential impact of different task experiences on new tasks, which easily leads to excessive borrowing of low-relevance experience, thus resulting in a decrease in recognition rate. In addition, since there are often large differences between source domain data and target domain data, the pre-training model needs to have good generalization performance and should be able to effectively extract the common features of source domain target data; for the meta-learning method combining the base learner and the meta-learner, the loss function of the base learner is easy to converge to the local minimum, and the commonly used cross entropy loss function will cause the model to obtain overconfident classification results, thereby making the base learner model in the pre-training stage lack the necessary generalization performance. In addition, in the design of the existing meta-learning method combining the base learner and the meta-learner, in order to simplify the difficulty of model design, the same loss function is often used in the meta-training task recognition training model and the meta-test task recognition training model, while ignoring the potential differences between the source domain data corresponding to the meta-training task and the target domain data corresponding to the meta-test task, and failing to design differentiated loss functions for different recognition tasks of the meta-training task and the meta-test task, which reduces the model transfer efficiency and the target domain recognition accuracy. In fact, in the pre-training stage corresponding to the meta-training task, the design of the loss function in the recognition model should focus on the generalization performance of the model and focus on extracting the common features of the source domain target data; under small sample conditions, there is often an imbalance in the number of samples between multiple target categories to be identified. In the back-propagation process, the category with more samples accounts for a higher proportion of the loss function, which can easily guide the model to optimize in the direction that is conducive to outputting the classification results of this category, thereby reducing the recognition rate of targets in the category with fewer samples. Therefore, in the fine-tuning stage corresponding to the meta-test task, the design of the loss function in the recognition model should avoid the problem of decreased recognition rate of targets in the category with fewer samples caused by the imbalance of the number of samples of different categories, and focus on reducing the weight of the easy-to-classify category in the loss function to balance the proportion of different categories in the loss function, focusing on extracting the individual characteristics of small sample target data and enhancing the matching degree of the meta-test task recognition model to the recognition task.

[0006] In order to solve the problem of poor HRRP multi-category recognition accuracy of radar targets under small sample conditions, how to design targeted models in the pre-training and fine-tuning stages is an important issue. In the pre-training stage of the meta-training task, how to design a basic learner loss function with strong generalization ability, and then conduct targeted transfer and reference of different task learning experience in the meta-learner; in the fine-tuning stage of the meta-test task, how to design a recognition model loss function that can effectively extract the individual characteristics of small sample target data; the two stages work together to use the meta-training task experience to guide the task to be identified to perform effective small sample recognition and improve the recognition accuracy of the target to be identified. This is one of the difficulties in the field of multi-category small sample intelligent recognition of radar targets. III. Summary of the invention

[0007] 1. Technical problems to be solved

[0008] In order to solve the problem of poor HRRP multi-category recognition accuracy of radar targets under small sample conditions, how to build a suitable meta-learning recognition framework, decouple the characteristics of a single task from the commonalities between tasks, design suitable basic learners and meta-learners to transfer experience between the characteristics of a single task and the commonalities between tasks, and design targeted recognition model loss functions in the pre-training stage and fine-tuning stage to achieve effective small sample recognition of radar targets; for the construction of basic learners, how to design a loss function with strong generalization ability to avoid the defects of the commonly used cross entropy loss function that easily leads to overconfident classification results and easy convergence to local minimum values. Then, the common features of the source domain target data can be effectively extracted to reserve the necessary universal experience to be transferred for the subsequent meta-test tasks; for the construction of the meta-learner, how to achieve targeted transfer and reference of learning experience of different tasks to guide the new recognition task to perform effective small sample recognition and improve the recognition accuracy of small sample targets of the new task; for the construction of the recognition model of the meta-test task, how to design a suitable loss function to avoid the problem of decreased recognition rate of targets in the category with fewer samples caused by the imbalance of the number of samples in different categories, effectively extract the individual features of the small sample target data, and enhance the matching degree of the meta-test task recognition model to the recognition task.

[0009] 2. Technical solution

[0010] The target intelligent recognition method based on loss fuzzy correction of the present invention comprises the following steps:

[0011] Step 1: Constructing the data set for deep meta-learning. Based on the radar high-resolution range image data of N types of targets to be identified, construct a small sample training data set for the meta-test task and a test data set for the meta-test task, and classify the tasks to be identified into the meta-test task set. Based on the accumulated data of M types of non-targets to be identified that are of different categories or different data sources from the target data to be identified, construct a meta-training task set by randomly sampling N types of target data, and use all the N types of non-target data to be identified as the training data set for the meta-training task.

[0012] The meta-learning method specifically includes a neural network adaptation method, a metric learning adaptation method, a basic learner and a meta-learner adaptation method, and a Bayesian meta-learning adaptation method. Among them, the combination of the basic learner and the meta-learner can decouple the task characteristic modeling from the commonality modeling between tasks. Through the mutual communication between the two, the model can achieve the best in both task characteristics and commonality, thereby improving the generalization ability of the model while maintaining the accuracy of the model. The data structure of traditional machine learning is generally divided into a training set and a test set. Both the training set and the test set need to contain a large amount of sample data. Deep meta-learning is similar to traditional machine learning and is also divided into a meta-training set and a meta-test set. The difference is that the meta-training set and the meta-test set are not sample data, but a collection of each task, and each task contains a corresponding training data set and a test data set. Unlike traditional machine learning and classical meta-learning, the data structure of the meta-learning model of the present invention is divided into a meta-training task set and a meta-test task set. The meta-test task set contains corresponding training data sets and test data sets, and in order to improve data utilization, the meta-training task set only contains corresponding training data sets.

[0013] For the recognition task of N types of radar targets, based on the radar high-resolution range image data of the N types of targets to be recognized, a meta-test task small sample training data set and a meta-test task test data set are constructed, that is, the meta-test task training data set is a small sample data set, and the task to be recognized is classified as a meta-test task set; based on M types of non-to-be-recognized targets of different target categories or different data sources from the target data to be recognized, data are accumulated, where M≥N, and a meta-training task set is constructed by randomly extracting N types of target data from the M types. Each meta-training task contains N types of non-to-be-recognized targets, and each meta-training task only contains the meta-training task training data set.

[0014] Different from the classic meta-learning method which divides the meta-training task data into training data set and test data set, the present invention uses all the data in the meta-training task as the meta-training task training data set, thereby improving the training utilization rate of the meta-training task data and the calculation accuracy and stability of the characteristic parameters and loss values ​​of a single meta-training task, which is beneficial to provide more reliable task learning experience for subsequent meta-learners, and is used to guide the precise design of small sample classifiers for targets to be identified.

[0015] It should be emphasized that, in view of the small sample characteristics of the target data to be identified, compared with the radar high-resolution range image data of the N types of targets to be identified, the differences of the accumulated data of the non-to-be-identified targets of the present invention can be reflected in two aspects: either reflected in different target types (for example, the target to be identified is a ship target, and the accumulated data of the non-to-be-identified targets involves aircraft targets); or reflected in different data types (for example, for both being ship targets, the radar high-resolution range image data of the target to be identified is measured data, while the accumulated data of the non-to-be-identified targets is simulated data); or both of the above aspects are different (for example, the radar high-resolution range image data of the target to be identified is measured data, while the non-to-be-identified targets are ground vehicle targets, and their accumulated data are one-dimensional radial sampling data of infrared images or radar high-resolution range image simulation data).

[0016] Step 2: Meta-learning model design; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss base learner and a task loss inverse weighted update meta-learner; the fuzzy truncated cosine loss base learner uses the fuzzy truncated cosine meta-training loss function as the convolutional neural network loss function of the meta-training task; the task loss inverse weighted update meta-learner uses the inverse of the meta-training task loss value to perform weighted correction and update on the meta-learner parameters based on the smoothed mean absolute error loss function; the input of the meta-learning model is the meta-training task set, and the output of the meta-learning model is the updated model parameters;

[0017] The meta-learning model of the present invention is designed as a combination of a fuzzy truncated cosine loss base learner and a task loss inverse weighted update meta-learner. The base learner models the task characteristics, requiring the base learner to be able to discover the inherent laws of each task data set. The deep learning model relies on a strong representation ability, can extract deep-level features of the target, and has a good degree of fit to the data. The convolutional neural network is a typical representative of the deep learning model, which uses a convolutional layer to automatically extract radar target data features and constrains the features through a loss function. When the number of target data to be identified is sufficient, the model can be directly trained from scratch, and the separability of the target features can be further improved through the loss function. When the number of HRRP samples of the target to be identified is insufficient, training from scratch will lead to the appearance of underfitting of the model. To solve this problem, the model can be pre-trained using other target measured data sets or simulation data sets that are unrelated to the target data, that is, the model is first pre-trained using the meta-training task, and then the pre-trained model is fine-tuned using the target data to be identified corresponding to the training data set in the meta-test task. Therefore, the pre-trained model needs to have good generalization performance, that is, the model has the ability to extract common features of the target data.

[0018] Aiming at the radar target HRRP multi-category small sample recognition task, the present invention designs a fuzzy truncated cosine loss based learner, uses a one-dimensional convolutional neural network (CNN) as the learning and training model, and constructs a fuzzy truncated cosine element training loss function as the loss function of the one-dimensional CNN. The fuzzy truncated cosine loss based learner includes an input layer, three convolutional layers (Conv), three pooling layers (POOL), a batch normalization layer (BN), two fully connected layers (FC) and an output layer. The specific connection method is:

[0019] Conv1→POOL1→Conv2→POOL2→Conv3+BN→POOL3→FC1→FC2

[0020] Among them, the input layer imports a small batch of data in the training data set of each meta-training task; the first two convolutional layers Conv1 and Conv2 are standard convolutional layers; the third convolutional layer Conv3 is a standardized convolutional layer, and a batch normalization layer BN is added on the basis of the convolutional layer Conv3 to enhance the gradient change of the fuzzy truncated cosine meta-training loss function. The three convolutional layers all use one-dimensional convolution kernels with sizes of A1, A2 and A3, numbers of B1, B2 and B3, step length of D, and the filling method is the same filling; the pooling layer POOL is an average pooling layer with a step length of H; the number of neurons in the first fully connected layer FC1 is F, and the number of neurons in the second fully connected layer FC2 is equal to the number of target categories to be identified N; the output layer uses the softmax function to calculate the category probability of each sample, and calculates the loss function value based on the calculation results.

[0021] It is noted that the HRRP of radar targets is sensitive to attitude angles, and the HRRP waveforms of the same target with different attitude angles are quite different; the HRRP samples corresponding to some attitude angles contain more scattering point information and are easy to identify, while the HRRP samples corresponding to other attitude angles contain less scattering point information and are more difficult to identify. However, HRRP samples of all attitude angles are equally important for target recognition and determine the generalization performance of the model. In order to ensure that the pre-trained model can well extract the complete attitude angle invariant features of HRRP, it is necessary to increase the output probability of the category corresponding to the HRRP samples that are more difficult to identify. The classic cross entropy loss function is difficult to meet this requirement. In response to the above problems, the present invention designs a fuzzy truncated cosine element training loss function as the loss function of the one-dimensional CNN in the basic learner. The fuzzy truncated cosine element training loss function includes two main parts. The first part is the fuzzy cosine loss function L F , which is used to solve the problem of overconfidence in the model classification results. By fuzzifying the output results, the gap between the outputs of each neuron is narrowed, so that the output results of each neuron play a role in the propagation process, avoiding the phenomenon of overconfidence in the model. The second part is the truncated cosine loss function LT , which is used to solve the problem of low output probability of some HRRP corresponding categories that are difficult to identify. By using the truncation characteristics of the truncation function, only the output results that meet the conditions are back-propagated, and by equivalently increasing the output result weight of this part of the sample data, the model can better extract the features of easily confused targets.

[0022] The fuzzy truncated cosine meta-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 of each meta-training task can be expressed as:

[0023]

[0024] Among them, the fuzzy cosine loss function L F for

[0025]

[0026] Truncated cosine loss function L T for

[0027]

[0028] In the above three formulas, m is the number of sample data in the gradient update process; x i Represents the fully connected layer output feature vector corresponding to the i-th sample; y i is the true label of the i-th sample; represents the output result of the i-th sample in the output layer; W j is the jth column of the weight matrix W of the fully connected layer, which represents the weight vector corresponding to the jth target; log2(·) represents the logarithmic function with base 2; a, β, s and α are positive real parameters; among them, the positive parameter a satisfies 0≤a≤1-cos(2π / N), and its role is to enhance the feature vector x i With the y i The weight vector corresponding to the class target The angle constraint of ; 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 0≤β≤1 is the fuzzy weight, which is used to adjust the output weight of the fuzzy loss component; the real parameter α is used to adjust the weight of the truncated cosine loss function; φ (·,·) Represents the angle between two vectors of the same dimension; Represents the truncation function.

[0029] Using the truncation function It is possible to perform back propagation only on the output results that meet the conditions, by appropriately increasing the truncated cosine loss function L TThe weight α can make the model better extract the features of easily confused targets; among them, the truncation function It can be expressed as:

[0030]

[0031] Where T represents the cutoff threshold, which satisfies 0.5≤T≤1; u(·) represents the modified step function, which can be expressed as

[0032]

[0033] Where x represents the independent variable of the modified step function. Note that L t Only if the output result satisfies , can participate in back propagation. By equivalently increasing the output weight of the corresponding sample data, the model can better extract the features of easily confused targets.

[0034] The fuzzy truncated cosine element training loss function is based on the large margin cosine loss. While constraining the inter-class features, it can continuously reduce the angular distance of the intra-class features and improve the separability of the features. If the angle between the two vectors x and y is φ (x,y) , then the cosine distance between two vectors x and y is the cosine value of the angle between them, which can be expressed as the inner product of the normalized vectors cos(φ (x,y) )=x T y / (||x||||y||), the superscript T represents transpose, ||·|| represents taking the Euclidean norm, and the cosine distance can effectively avoid the adverse effect of the module values ​​of different vectors on the distance measurement. is the weight vector W j and the eigenvector x i The similarity measure between them is that the cosine of the angle is The interval is monotonically decreasing. The smaller the angle between the vectors, the larger the cosine value. The eigenvector x i The greater the probability that the corresponding i-th sample belongs to category j, the greater the probability that the corresponding i-th sample belongs to category j. In order to further increase the inter-class distance of target features of different categories, the present invention introduces a hyperparameter a. When When , the i-th sample is determined to belong to category j. The role of the hyperparameter a is to enhance the feature vector x i With weight vector Angle constraint.

[0035] The meta-learner in the meta-learning model design mainly models the commonalities between tasks. Its function is to summarize the training experience on all tasks after each basic learner training is completed, synthesize new experience, and feed it back to the basic learner. The meta-learner comes in various forms and can be any model based on stochastic gradient descent or a parameter update algorithm. In order to reduce the computational complexity and improve the accuracy and efficiency of parameter updates, the method of the present invention designs a task loss inverse weighted update meta-learner. Based on the smoothed mean absolute error loss function, the inverse of the meta-training task loss value is used to perform weighted correction and update on the meta-learner parameters. By fully considering the differential impact of different task experiences on new tasks, the learning experience of different meta-training tasks is specifically referenced, avoiding the problem of reduced recognition rate due to excessive reference to low-relevance experience, so that the initialization model parameter θ of the meta-learner is M Update towards the common direction of highly related tasks.

[0036] The main purpose of meta-learning is to find the common direction of highly related tasks so that the initialization model parameters θ of the meta-learner M Close to the characteristic parameters of the highly relevant meta-training tasks in the base learner, that is, the parameter θ M Minimize the error distance between all meta-training task characteristic parameters, that is, the learning objective function of the meta-learner is as follows

[0037]

[0038] Among them, D k is the characteristic parameter θ of the kth meta-training task Bk and the meta-learner parameters θ M Different from the commonly used meta-learner objective function design, the inequality measure is ignored in formula (6), and only the main error term is retained. On the basis of ensuring the accuracy of parameter update, the difficulty of subsequent gradient solution and the computational complexity of meta-learner parameter update are reduced, and the parameter update efficiency is improved.

[0039] In order to prevent the gradient explosion problem of the mean square error (corresponding to the Euclidean space distance) and avoid the non-smooth gradient defect of the mean absolute error at zero, the method of the present invention constructs a smooth mean absolute error loss function. k Specifically expressed as

[0040]

[0041] The learning goal of the meta-learner is to minimize the error between the parameters of the two learners. According to Lagrange’s theorem, D can be calculated k In the parameter θ M The gradient at D is minimized k For ||θBk -θ M ||<1, for distance D k Taking the derivative we get:

[0042]

[0043] Meta-learners often assume that each meta-training task has an effect on the meta-learner parameters θ M The influence of each task on the meta-learner parameter θ is the same. M The influence of is indifferently superimposed to calculate the common direction of all tasks. Considering the differential influence of different meta-training task experiences on new tasks, the task loss value L of the kth meta-training task is k It can reflect the degree of difference in learning experience to a certain extent. The smaller the loss value, the larger its reciprocal, which means the better the training effect, that is, the learning experience of this task is more worthy of reference; on the contrary, the larger the loss value, the smaller its reciprocal, which means the worse the learning effect of the meta-training task, and the lower the reference degree of learning experience. Therefore, in order to avoid excessive reference to low-relevance experience and make full use of the bias influence of different meta-training tasks on the parameter update of the meta-learner, the method of the present invention uses the reciprocal of the loss value of the meta-training task to perform weighted correction on the parameter update process of the meta-learner, so that the meta-learner can draw on the experience of each meta-training task in a targeted manner, reduce excessive reference to low-relevance experience, and improve the degree of reference to high-relevance experience. The parameter update method of the meta-learner weighted by the reciprocal of the task loss is as follows:

[0044]

[0045] in, is the updated parameter of the meta-learner, ε0 is the update step of the meta-learner; the task loss value L k The reciprocal of L k The change of is directly negatively correlated, and the weighted update method is intuitive and concise. Considering the function convergence problem in the parameter update process, min(·,·) means taking the smaller value of the two parameters, which is used to prevent L k If the value is too small, the parameter update process will not converge.

[0046] Step 3 randomly selects K meta-training tasks from the meta-training task set to train the basic learner; for the kth meta-training task, the meta-training task training data set is used to perform batch training on the fuzzy truncated cosine loss basic learner, and the characteristic parameter θ of the kth meta-training task is obtained through iterative update optimization of the convolutional neural network. Bk And the loss value L k ; Pass the characteristic parameters and loss values ​​of the K meta-training tasks forward to the task loss inverse weighted update meta-learner, and go to step 4;

[0047] K meta-training tasks are randomly selected from the meta-training task set to train the basic learner. For the kth meta-training task, the fuzzy truncated cosine loss basic learner designed in step 2 is used to train in batches in combination with the training data set of the kth meta-training task. According to the fuzzy truncated cosine meta-training loss function of formula (1), the convolutional neural network is used to iteratively update and optimize the training to obtain the characteristic parameter θ of the kth meta-training task. Bk and the corresponding task loss value L k ; After all K meta-training tasks have completed the basic learner training, the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK} and the corresponding task loss value {L B1 ,L B2 ,…,L Bk ,…,L BK}Forward pass to the task loss inverse weighted update meta-learner, and go to step 4 to update the meta-learner parameters.

[0048] Step 4: Update the meta-learner with the inverse of the task loss based on the characteristic parameters of the K meta-training tasks, and use the inverse of the corresponding meta-training task loss value to update the meta-learner parameter θ M Perform weighted correction update and update the updated parameters Feedback to the fuzzy truncated cosine loss based learner; To initialize the model parameters of the fuzzy truncated cosine loss based learner, re-execute step 3 for a new round of optimization training;

[0049] The inverse of the task loss weighted update meta-learner is based on the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK}, using the corresponding meta-training task loss value {L B1 ,L B2 ,…,L Bk ,…,L BK}, based on formula (9) for the meta-learner parameter θ M Perform weighted correction update and update the updated parameters Feedback to the fuzzy truncated cosine loss based learner; To initialize the model parameters of the fuzzy truncated cosine loss based learner, re-execute step 3 for a new round of optimization training;

[0050] Step 5: Repeat steps 3 and 4 until N is reached. cAt the end of the cycle, save the inverse of the last task loss and weight the update meta-learner parameters A convolutional neural network recognition model with the same structure as the fuzzy truncated cosine loss base learner is constructed, and a multi-class balanced cosine angle constraint loss function is used as the convolutional neural network loss function of the meta-test task; As the initialization parameters of the convolutional neural network recognition model, the small sample training data set of the meta-test task is imported in batches to train and update the recognition model. The recognition model training is terminated when the multi-class balanced cosine angle constraint loss function of the convolutional neural network converges and no longer decreases. The trained convolutional neural network recognition model is used to classify and recognize the meta-test task test data set, and the classification and recognition accuracy of the N types of targets to be identified is evaluated.

[0051] In the fine-tuning stage corresponding to the meta-test task, the accuracy of target recognition under small sample conditions can be greatly improved by fine-tuning the pre-trained model with high generalization performance. Under small sample conditions, there is often an imbalance in the number of samples between multiple categories. In the back-propagation process, the category with more samples accounts for a higher proportion of the loss function, which easily guides the model to optimize in the direction that is conducive to outputting the classification results of this category, thereby reducing the recognition rate of targets in the category with fewer samples. To this end, it is necessary to design the loss function of the convolutional neural network recognition model in the fine-tuning stage corresponding to the meta-test task in a targeted manner.

[0052] While the large margin cosine loss constrains the inter-class features, it can continuously reduce the angular distance of the intra-class features, thereby improving the separability of the features. To this end, the present invention constructs a convolutional neural network recognition model with the same structure as the fuzzy truncated cosine loss basic learner as a fine-tuning model corresponding to the meta-test task, and adopts a multi-class balanced cosine angle constraint loss function as the convolutional neural network loss function of the meta-test task; the multi-class balanced cosine angle constraint loss function calculates the Focal loss based on the cosine distance, and uses the enhanced marginal cosine loss to constrain the inter-class features while continuously reducing the angular distance of the intra-class features, thereby improving the separability of the features; at the same time, the loss function reduces the weight of the easily distinguishable category in the loss function by introducing the Focal loss, thereby balancing the proportion of each category in the loss function, thereby avoiding the problem of reduced recognition rate of targets in categories with fewer samples due to the imbalance of the number of samples in different categories.

[0053] In the convolutional neural network recognition model in the fine-tuning stage corresponding to the meta-test task, the multi-class balanced cosine angle constraint loss function can be expressed as:

[0054]

[0055] The parameter γ is used to adjust the output weight, satisfying 0≤γ≤5.

[0056] The meta-learning training phase is structurally divided into two loop structures, the inner loop refers to the cyclic update of parameters between the base learner and the meta-learner, and the outer loop refers to the training round cycle of the inner loop. First, the model parameters of the base learner are initialized, and the recognition tasks are extracted from the meta-training set and imported into the base learner for learning. The base learner parameters are updated according to the fuzzy truncated cosine meta-training loss function; secondly, the model parameters learned by the base learner are imported into the meta-learner for parameter update by weighted update of the inverse of the task loss, and the updated parameters are fed back to the base learner for re-assignment of parameters, and a new round of meta-training task extraction and base learner training is carried out. The cycle is repeated until the outer loop number is reached and the final CNN model parameters are saved.

[0057] The meta-learning test phase uses the training parameters of the meta-learning training phase to initialize the parameters of the recognition model, so as to transfer the previous learning experience to the meta-test set, and use the initialized recognition model for classification and recognition to verify the effectiveness of the method. First, the CNN model parameters saved in the meta-learning training phase are used as the initial parameters of the recognition model, and the small sample training data set in the meta-test set is imported to train the recognition model until the multi-class balanced cosine angle constraint loss function converges, and the meta-test set training is terminated; then, the recognition model and parameters obtained from the meta-test set training are used to evaluate the classification and recognition of the test data set of the meta-test set, and the classification and recognition accuracy of each type of target in the test data set is calculated.

[0058] 3. Beneficial effects

[0059] Compared with the background technology, the present invention has the following beneficial effects:

[0060] 1) To address the problem of poor multi-category recognition accuracy of radar target HRRP under small sample conditions, a meta-learning recognition framework is constructed by combining a fuzzy truncated cosine loss base learner with a task loss inverse weighted update meta-learner. This framework decouples the characteristics of a single task from the commonalities between tasks, and performs experience transfer on the characteristics of a single task and the commonalities between tasks, thus achieving effective recognition of radar target HRRP with small samples.

[0061] 2) Fully consider the potential differences between the source domain data corresponding to the meta-training task and the target domain data corresponding to the meta-testing task, and carry out targeted loss function design of the recognition model in the pre-training stage of the meta-training task and the fine-tuning stage of the meta-testing task respectively; in the loss function design of the recognition model in the pre-training stage, we pay attention to the generalization performance of the model and focus on extracting the common features of the source domain target data; in the loss function design of the recognition model in the fine-tuning stage, we pay attention to balancing the proportion of different categories in the loss function and focus on extracting the individual features of the small sample target data to be identified; thus, we improve the model transfer efficiency and the target domain recognition accuracy.

[0062] 3) For the CNN construction of the basic learner, a fuzzy truncated cosine meta-training loss function is designed. On the one hand, the fuzzy cosine loss function is constructed to solve the problem of overconfidence in the model classification results. By fuzzifying the output results, the gap between the outputs of each neuron is narrowed, so that the output results of each neuron play a role in the propagation process, avoiding the phenomenon of overconfidence of the model; on the other hand, the truncated cosine loss function is constructed to solve the problem of low output probability of the corresponding category of some sample data that are difficult to recognize. By using the truncation characteristics of the truncation function, only the output results that meet the conditions are back-propagated, and the output weights of this part of the sample data are equivalently increased, so that the model can better extract the features of easily confused targets; the synergistic effect of the two aspects improves the generalization ability of the fuzzy truncated cosine meta-training loss function, thereby effectively extracting the common features of the source domain target data, and reserving the necessary general experience to be transferred for subsequent meta-testing tasks.

[0063] 4) A meta-learner learning objective function containing only the main error term is constructed. By ignoring the inequality metrics with low influence, the difficulty of subsequent gradient solution and the computational complexity of meta-learner parameter update are reduced while ensuring the accuracy of parameter update, thereby improving the efficiency of parameter update.

[0064] 5) A smoothed mean absolute error loss function of the meta-learner is constructed, which avoids the gradient explosion problem of the mean square error, overcomes the non-smooth gradient defect of the mean absolute error at zero, improves the modeling accuracy of the error term of the meta-learner objective function, and provides a model guarantee for the efficient update of the subsequent meta-learner parameters.

[0065] 6) A weighted update method based on the inverse of the task loss of the meta-learner parameters is designed. The negative correlation between the meta-training task loss value and the task training effect is used to avoid excessive reference to low-relevance experience and increase the degree of reference to high-relevance experience. By taking the smaller operation, the problem of non-convergence of the parameter update process caused by the meta-training task loss value being too small is avoided.

[0066] 7) For the fine-tuning CNN model of the meta-test task, a multi-class balanced cosine angle constraint loss function is designed. The enhanced marginal cosine loss is used to constrain the inter-class features, while the angular distance of the intra-class features is continuously reduced to improve the separability of the features. At the same time, the Focal loss is introduced to reduce the weight of the easy-to-distinguish categories in the loss function, so as to balance the proportion of each category in the loss function, avoid the problem of decreased recognition rate of targets in categories with fewer samples due to the imbalance of the number of samples in different categories, and enhance the matching degree of the meta-test task recognition model to the recognition task. IV. Description of the drawings

[0067] Figure 1 It is a schematic diagram of the meta-learning data structure of the method of the present invention;

[0068] Figure 2 It is a schematic diagram of the structure of the meta-learning model of the method of the present invention;

[0069] Figure 3 It is a schematic diagram of the structure of the basic learner of the method of the present invention;

[0070] Figure 4 This is a flow chart of the target intelligent recognition method based on loss fuzzy correction. V. Specific implementation methods

[0071] The present invention is further described below in conjunction with the accompanying drawings. The embodiments of the present invention are used to explain the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modifications and changes made to the present invention fall within the protection scope of the present invention.

[0072] To verify the effectiveness of the method described in the present invention, this specific implementation method provides two examples. The first example is for the HRRP recognition application of multiple types of ship target radars, and the second example is for the HRRP recognition application of multiple types of aircraft target radars.

[0073] Embodiment 1:

[0074] Example 1 is an application of radar HRRP recognition of five types of ship targets, which can be specifically divided into the following steps:

[0075] Step A-1: ​​Construct a deep meta-learning dataset, whose meta-learning data structure is as follows Figure 1 As shown in the figure; firstly, the simulation data of HRRP of M=10 types of ship targets are obtained, and the meta-training task set is constructed by randomly extracting N=5 types of target data, and all the simulation data of N types of non-to-be-identified ship targets are used as the meta-training task training data set; in view of the non-cooperation of ship targets in actual experiments, only a small number of small sample HRRP measured data of ship targets to be identified can be obtained. According to the radar measured HRRP data of N types of ship targets to be identified, the small sample training data set of the meta-test task and the test data set of the meta-test task are constructed, and the task to be identified is classified into the meta-test task set.

[0076] Step A-2: Design a meta-learning model; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss base learner and a task loss inverse weighted update meta-learner. Its model structure is as follows: Figure 2 As shown. The fuzzy truncated cosine loss base learner uses the fuzzy truncated cosine meta-training loss function as the convolutional neural network loss function of the meta-training task; the task loss inverse weighted update meta-learner is based on the smoothed mean absolute error loss function, and uses the inverse of the meta-training task loss value to perform weighted correction and update on the meta-learner parameters; the input of the meta-learning model is the meta-training task set, and the output of the meta-learning model is the updated model parameters;

[0077] Step A-3: extract 5 types of target data from the simulation data of HRRP of 10 types of ship targets without duplication to form a set of 252 meta-training tasks, and randomly extract K=10 meta-training tasks from the meta-training task set to train the basic learner;

[0078] CNN model parameters θ for the fuzzy truncated cosine loss based learner B Initialize, including the weights of the convolutional layer {w conv1 ,b conv1 ,w conv2 ,b conv2 , w conv3 ,b conv3}, the weights of the fully connected layer {w fc ,b fc ,w fc2 ,b fc2 The structural parameters of the fuzzy truncated cosine loss base learner are: N = 5, A1 = 9, A2 = 9, A3 = 9, B1 = 8, B2 = 16, B3 = 32, D = 1, H = 2, F = 100; the specific design of the corresponding fuzzy truncated cosine loss base learner is as follows: Figure 3 shown.

[0079] For the kth meta-training task, the fuzzy truncated cosine loss based learner is used to perform batch training in combination with the training data set of the kth meta-training task. According to the convolutional neural network fuzzy truncated cosine meta-training loss function of formula (1), the Adam optimizer is used to iteratively optimize and update the parameters, and then the characteristic parameter θ of the kth meta-training task is obtained. Bk and the corresponding task loss value L k ; After all K meta-training tasks have completed the basic learner training, the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK} and the corresponding task loss value {L B1 ,L B2 ,…,L Bk ,…,L BK}Forward pass to the task loss inverse weighted update meta-learner, and go to step A-4 to update the meta-learner parameters.

[0080] Step A-4: Update the meta-learner based on the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK}, using the corresponding meta-training task loss value {L B1,L B2 ,…,L Bk ,…,L BK}, based on formula (9) for the meta-learner parameter θ M Perform weighted correction update and update the updated parameters Feedback to the fuzzy truncated cosine loss base learner in step A-3; Re-initialize the model parameters of the fuzzy truncated cosine loss based learner and re-execute step A-3 for a new round of optimization training;

[0081] Step A-5: The process of the target intelligent recognition method based on loss blur correction is as follows Figure 4 According to the processing flow, step A-3 and step A-4 are executed repeatedly until N is reached. c = 25 cycles are completed and the final meta-learner parameters are saved Based on the fuzzy truncated cosine loss basic learner, a CNN recognition model with the same structure is constructed. As the initialization parameters of the CNN recognition model, the small sample training data set of the meta-test task is imported in batches to perform model training and update. The model training is terminated when the CNN multi-class balanced cosine angle constraint loss function expressed by equation (10) converges and no longer decreases. The trained CNN recognition model is used to classify and recognize the meta-test task test data set, and the classification and recognition accuracy of the measured data of the five types of ship targets to be identified is evaluated.

[0082] The experimental analysis results show that compared with the CNN model, the deep transfer learning model (DTL) and the meta-learning model (UBML) with unbiased reference to task experience, the meta-learning model based on loss fuzzy correction designed by the present invention greatly improves the recognition performance of the measured HRRP of multiple types of ship targets under small sample conditions by differentially and targetedly transferring and utilizing the previous experience of different simulated ship target recognition tasks. Under small sample conditions, the CNN model is directly trained and tested from beginning to end. Due to the reduction of training data, the CNN network is prone to overfitting and poor recognition rate; DTL can only transfer single task characteristics, and its recognition is limited by the experience quality of a single source task, and the recognition effect of small samples is poor; UBML indiscriminately references the learning experience of the meta-training task, and can perform more accurate classification and recognition through the transfer of common experience of the task. However, the same CNN loss function is used in the pre-training stage and the fine-tuning stage, which makes it difficult to extract the general features of the source domain target data in the pre-training stage while extracting the individual features of the small sample target data to be identified in the fine-tuning stage. In addition, excessive reference to low-relevance experience leads to poor recognition rate of small samples of ship targets, and it is difficult to avoid the imbalance of the number of samples of different categories of targets. The recognition rate of targets with fewer categories decreases; the meta-learning model based on loss fuzzy correction designed by the present invention performs a weighted correction of the inverse of the task loss on the learning experience of different tasks in the meta-learner, guides the new recognition task to perform effective small sample recognition, and uses the fuzzy truncated cosine meta-training loss function of the basic learner CNN model to extract the common features of the source domain target data. The meta-test task is used to fine-tune the multi-class balanced cosine angle constraint loss function of the CNN model, extracts the individual features of the small sample target data to be identified, balances the proportion of each category in the loss function, avoids the problem of reduced recognition rate of targets with fewer samples caused by the imbalance of the number of samples of different categories, enhances the matching degree of the meta-test task recognition model to the recognition task, and improves the comprehensive migration efficiency of the model and the target domain recognition accuracy. In summary, the method of the present invention is better than other existing learning models as a whole, and is more suitable for the classification and recognition of small sample ship targets than previous transfer learning methods, which reflects the superiority of the method of the present invention.

[0083] Embodiment 2:

[0084] Example 2 is an application of radar HRRP recognition of four types of aircraft targets, which can be specifically divided into the following steps:

[0085] Step B-1: Construct a deep meta-learning dataset, whose meta-learning data structure is as follows Figure 1As shown in the figure. Due to the lack of simulation models for aircraft targets, the accumulated measured HRRP data of ground vehicle targets can be used to construct meta-learning training tasks. First, the measured HRRP data of M = 8 types of vehicle targets are obtained, and the meta-training task set is constructed by randomly extracting N = 4 types of target data, and all the measured data of N types of non-identified vehicle targets are used as the meta-training task training data set; in view of the non-cooperativeness of actual test aircraft targets, only a small number of small sample HRRP measured data of aircraft targets to be identified can be obtained. Based on the radar measured HRRP data of N types of aircraft targets to be identified, a small sample training data set of the meta-test task and a test data set of the meta-test task are constructed, and the task to be identified is classified as the meta-test task set.

[0086] Step B-2: Design a meta-learning model; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss base learner and a task loss inverse weighted update meta-learner. Its model structure is as follows: Figure 2 As shown. The fuzzy truncated cosine loss base learner uses the fuzzy truncated cosine meta-training loss function as the convolutional neural network loss function of the meta-training task; the task loss inverse weighted update meta-learner is based on the smoothed mean absolute error loss function, and uses the inverse of the meta-training task loss value to perform weighted correction and update on the meta-learner parameters; the input of the meta-learning model is the meta-training task set, and the output of the meta-learning model is the updated model parameters;

[0087] Step B-3: extract 4 types of target data from the measured data of HRRP of 8 types of vehicle targets without duplication to form a set of 70 meta-training tasks, and randomly extract K=10 meta-training tasks from the meta-training task set to train the basic learner;

[0088] CNN model parameters θ for the fuzzy truncated cosine loss based learner B Initialize, including the weights of the convolutional layer {w conv1 ,b conv1 ,w conv2 ,b conv2 , w conv3 ,b conv3}, the weights of the fully connected layer {w fc ,b fc ,w fc2 ,b fc2 The structural parameters of the fuzzy truncated cosine loss base learner are: N = 5, A1 = 9, A2 = 9, A3 = 9, B1 = 8, B2 = 16, B3 = 32, D = 1, H = 2, F = 100; the specific design of the corresponding fuzzy truncated cosine loss base learner is as follows: Figure 3 shown.

[0089] For the kth meta-training task, the fuzzy truncated cosine loss based learner is used to perform batch training in combination with the training data set of the kth meta-training task. According to the convolutional neural network fuzzy truncated cosine meta-training loss function of formula (1), the Adam optimizer is used to iteratively optimize and update the parameters, and then the characteristic parameter θ of the kth meta-training task is obtained. Bk and the corresponding task loss value L k ; After all K meta-training tasks have completed the basic learner training, the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK} and the corresponding task loss value {L B1 ,L B2 ,…,L Bk ,…,L BK}Forward pass to the task loss inverse weighted update meta-learner, and go to step B-4 to update the meta-learner parameters.

[0090] Step B-4: Update the meta-learner based on the characteristic parameters {θ B1 ,θ B2 ,…,θ Bk ,…,θ BK}, using the corresponding meta-training task loss value {L B1 ,L B2 ,…,L Bk ,…,L BK}, based on formula (9) for the meta-learner parameter θ M Perform weighted correction update and update the updated parameters Feedback to the fuzzy truncated cosine loss base learner in step B-3; Re-execute step B-3 to perform a new round of optimization training for the newly initialized model parameters of the fuzzy truncated cosine loss based learner;

[0091] Step B-5: The process of the target intelligent recognition method based on loss blur correction is as follows Figure 4 According to the processing flow, step B-3 and step B-4 are executed repeatedly until N is reached. c =7 cycles end, save the final meta-learner parameters Based on the fuzzy truncated cosine loss basic learner, a CNN recognition model with the same structure is constructed. As the initialization parameters of the CNN recognition model, the small sample training data set of the meta-test task is imported in batches to perform model training and update. The model training is terminated when the CNN multi-class balanced cosine angle constraint loss function expressed by equation (10) converges and no longer decreases. The trained CNN recognition model is used to classify and recognize the meta-test task test data set, and the classification and recognition accuracy of the measured data of the four types of aircraft targets to be recognized is evaluated.

[0092] The experimental analysis results show that compared with the CNN model, the deep transfer learning model (DTL) and the meta-learning model (UBML) with unbiased reference to task experience, the meta-learning model based on loss fuzzy correction designed by the present invention has greatly improved the recognition performance of the measured HRRP of multiple types of aircraft targets under small sample conditions by differentially and targetedly transferring and utilizing the previous different measured vehicle target recognition task experiences. Under small sample conditions, the CNN model is directly trained and tested from beginning to end. Due to the reduction of training data, the CNN network is prone to overfitting and poor recognition rate; DTL can only transfer single-task characteristics, and its recognition is limited by the experience quality of a single source task, and the recognition effect of small samples is poor; UBML indiscriminately references the learning experience of the meta-training task, and can perform more accurate classification and recognition through the transfer of common experience of the task. However, the same CNN loss function is used in the pre-training stage and the fine-tuning stage, making it difficult to extract the common features of the source domain target data in the pre-training stage while extracting the individual features of the small sample target data to be identified in the fine-tuning stage. In addition, excessive reference to low-relevance experience leads to poor recognition rate of small samples of aircraft targets, and it is difficult to avoid the imbalance of the number of samples of different categories of targets. The recognition rate of targets with fewer categories decreases; the meta-learning model designed by the present invention based on loss fuzzy correction performs weighted correction of the inverse of task loss on the learning experience of different tasks in the meta-learner, guides the new recognition task to perform effective small sample recognition, and uses the fuzzy truncated cosine meta-training loss function of the basic learner CNN model to extract the common features of the source domain target data. The meta-test task is used to fine-tune the multi-class balanced cosine angle constraint loss function of the CNN model to extract the individual features of the small sample target data to be identified, balance the proportion of each category in the loss function, avoid the problem of reduced recognition rate of targets with fewer samples caused by the imbalance of the number of samples of different categories, enhance the matching degree of the meta-test task recognition model to the recognition task, and improve the comprehensive migration efficiency of the model and the target domain recognition accuracy. In summary, the method of the present invention is better than other existing learning models as a whole, and is more suitable for the classification and recognition of small sample aircraft targets than previous transfer learning methods, which reflects the superiority of the method of the present invention.

Claims

1. A target intelligent recognition method based on loss fuzzy correction, characterized in that: The following steps are involved: Step 1: Constructing the data set for deep meta-learning. Based on the radar high-resolution range image data of N types of targets to be identified, construct a small sample training data set for the meta-test task and a test data set for the meta-test task, and classify the tasks to be identified into the meta-test task set. Based on the accumulated data of M types of non-targets to be identified that are of different categories or different data sources from the target data to be identified, construct a meta-training task set by randomly sampling N types of target data, and use all the N types of non-target data to be identified as the training data set for the meta-training task. Step 2: Meta-learning model design; the meta-learning model is designed as a combination of a fuzzy truncated cosine loss base learner and a task loss inverse weighted update meta-learner; the fuzzy truncated cosine loss base learner uses the fuzzy truncated cosine meta-training loss function as the convolutional neural network loss function of the meta-training task; the task loss inverse weighted update meta-learner is based on the smoothed mean absolute error loss function and uses the inverse of the meta-training task loss value to perform weighted correction and update on the meta-learner parameters; The input of the meta-learning model is the meta-training task set, and the output of the meta-learning model is the updated model parameters; In the construction of the convolutional neural network of the fuzzy truncated cosine loss base learner, a fuzzy truncated cosine meta-training loss function is designed, which is composed of a linear combination of the fuzzy cosine loss function and the truncated cosine loss function. The fuzzy truncated cosine meta-training loss function of each meta-training task is expressed as Among them, m is the number of sample data in the gradient update process; x i Represents the fully connected layer output feature vector corresponding to the i-th sample; y i is the true label of the i-th sample; represents the output result of the i-th sample in the output layer; W j is the weight vector corresponding to the j-th target; a, β, s and α are positive real parameters; φ (·,·) Represents the angle between two vectors of the same dimension; represents the truncation function, which can be expressed as Where T represents the cutoff threshold, which satisfies 0.5≤T≤1; u(·) represents the modified step function, which can be expressed as The main error term in the learning objective function of the task loss inverse weighted update meta-learner is designed as a smoothed mean absolute error loss function, and the characteristic parameter θ of the kth meta-training task is Bk and the meta-learner parameters θ M The error between k Expressed as Among them, ||·|| means taking the Euclidean norm; The parameter update method of the task loss inverse weighted update meta-learner is in, is the updated parameter of the meta-learner, and min(·,·) means taking the smaller value of the two parameters; Step 3 randomly selects K meta-training tasks from the meta-training task set to train the basic learner; for the kth meta-training task, the meta-training task training data set is used to perform batch training on the fuzzy truncated cosine loss basic learner, and the characteristic parameter θ of the kth meta-training task is obtained through iterative update optimization of the convolutional neural network. Bk And the loss value L k ; Pass the characteristic parameters and loss values ​​of the K meta-training tasks forward to the task loss inverse weighted update meta-learner, and go to step 4; Step 4: Update the meta-learner with the inverse of the task loss based on the characteristic parameters of the K meta-training tasks, and use the inverse of the corresponding meta-training task loss value to update the meta-learner parameter θ M Perform weighted correction update and update the updated parameters Feedback to the fuzzy truncated cosine loss based learner; To initialize the model parameters of the fuzzy truncated cosine loss based learner, re-execute step 3 for a new round of optimization training; Step 5: Repeat steps 3 and 4 until N is reached. c At the end of the cycle, save the inverse of the last task loss and weight the update meta-learner parameters A convolutional neural network recognition model with the same structure as the fuzzy truncated cosine loss base learner is constructed, and a multi-class balanced cosine angle constraint loss function is used as the convolutional neural network loss function of the meta-test task; As the initialization parameters of the convolutional neural network recognition model, the small sample training data set of the meta-test task is imported in batches to train and update the recognition model. The recognition model training is terminated when the multi-class balanced cosine angle constraint loss function of the convolutional neural network converges and no longer decreases. The trained convolutional neural network recognition model is used to classify and recognize the meta-test task test data set, and the classification and recognition accuracy of the N types of targets to be recognized is evaluated. The multi-class balanced cosine angle constraint loss function balances the proportion of each category in the loss function by reducing the weight of the easily distinguishable category in the loss function, which can be expressed as: The parameter γ is used to adjust the output weight, satisfying 0≤γ≤5.

Images