Hrrp identification model based on conditional convolution and correlation mapping and online updating method thereof

Abstract

The application discloses an HRRP identification model online updating method based on conditional convolution and correlation mapping, and comprises the following steps: obtaining a training sample set with a target model label and task dividing the training sample set according to the attitude angle range of the target; constructing an initial multi-layer convolutional neural network model and converting the initial multi-layer convolutional neural network model into a multi-layer convolutional neural network model with a conditional convolution structure through pre-training; expanding the expert convolution kernel, the routing module and the correlation mapping mask for the next new task; training the expanded model for the current new task; pruning the correlation mapping mask of the full connection classifier by using a model pruning method; obtaining another new task and re-executing until the training of each new task is completed; and identifying the target by using the trained model on the actual radar high-resolution range profile (HRRP). The conditional convolution layer is used to extract the features of the HRRP, the number of the expert convolution kernel of the conditional convolution layer is expanded, the model capacity is improved, and the model capacity can be fully improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar HRRP target recognition, and particularly relates to an HRRP recognition model online updating method based on conditional convolution and correlation mapping. BACKGROUND

[0002] In the actual engineering application of radar high resolution range profile (HRRP) target classification, due to the objective reasons that cooperative targets are not the main objects and non-cooperative targets are not easy to collect data, it is often difficult to establish a complete database to train a mature model. Therefore, the model is required to continuously learn using new samples in the use process. However, when the radar obtains new samples, if the new samples are directly used to update the trained parameters of the model, the model no longer performs good performance on the old task, and a catastrophic forgetting problem is caused. How to alleviate the catastrophic forgetting is a main difficulty in establishing a sustainable training model with excellent performance.

[0003] In the scene where the type of target is known, but the target attitude has not been fully acquired (we call it attitude expansion scene), the current existing methods are mostly based on regularization, dynamic expansion architecture method. For example, the classic continuous learning method is the paper "Overcoming catastrophic forgetting in neural networks" published by James Kirkpatrick et al. Its method is to make the elastic weight consolidation (EWC) based on regularization, he restricts the change of the weight which is more important to the past knowledge to resist the forgetting of knowledge, the paper "Continual Learning Through Synaptic Intelligence" published by Friedemann Zenke et al. is a kind of synaptic intelligence (SI), they are a kind of regularization strategy; The paper "Progressive Neural Networks" published by Andrei A. Rusu et al. is a kind of progressive neural network (PNN) based on dynamic architecture method, which learns new knowledge by assigning independent network structure to the data expansion of the new recognition task. When the model is tested, the model needs to provide the data of the current recognition which belongs to which network structure corresponding to the task; In addition, the paper "Lifelong Learning with Dynamically Expandable Networks" published by Jaehong Yoon et al. proposes a kind of dynamic expandable network (DEN), which is also based on the expansion of the model. These methods have alleviated the problem of catastrophic forgetting to some extent.

[0004] But this kind of method has a series of problems, among which the restriction of parameter modification of regularization method will lead to the compression of free parameter space with the growth of training task, which will lead to the problem of weakening of learning ability, while the dynamic expansion network will continuously rise with the expansion of network model, and the amount of calculation and model parameters, and this kind of dynamic expansion network needs task label, that is, to tell the model which task the data of the current recognition belongs to, and there is no method to combine the dynamic expansion network structure method which needs task label with radar HRRP target recognition.

[0005] In the prior art, a part of the model structure of the method is relatively simple, such as only using a multi-layer perception, and expanding the number of neurons in each layer to expand the recognition ability of a new task. The model capacity of this kind of method is too small, and the nonlinear representation ability is poor. Secondly, part of this kind of method needs the help of task labels and other auxiliary information to combine with a specific network structure to process tasks, but no method explicitly combines this kind of method with radar HRRP target recognition.

[0006] Secondly, another part of the method uses a regularization strategy to limit the variation range of network parameters to maintain the recognition ability. However, excessive limitation of the variation of the parameters will cause the network to have a catastrophic memory, that is, the learning ability of the network for new task data will decrease linearly. SUMMARY

[0007] In order to solve the above problems existing in the prior art, the present application provides an HRRP recognition model online updating method based on conditional convolution and correlation mapping. The technical problem to be solved by the present application is realized by the following technical scheme:

[0008] The present application provides an HRRP recognition model online updating method based on conditional convolution and correlation mapping, comprising:

[0009] S1: obtaining a training sample set with target model labels and dividing the training sample set according to the attitude angle range of the target;

[0010] S2: constructing an initial multi-layer convolutional neural network model and converting it into a multi-layer convolutional neural network model with a conditional convolution structure through pre-training, wherein the initial multi-layer convolutional neural network model comprises a feature extractor and a full-connection classifier with a mask structure;

[0011] S3: expanding the expert convolution kernel, the routing module and the correlation mapping mask of the multi-layer convolutional neural network model with the conditional convolution structure for the next new task;

[0012] S4: training the expanded multi-layer convolutional neural network model with the conditional convolution structure for the current new task;

[0013] S5: pruning the correlation mapping mask of the full-connection classifier using a model pruning method;

[0014] S6: performing performance recovery training on the pruned multi-layer convolutional neural network model using the current new task;

[0015] S7: obtaining another new task and re-executing steps S3 to S6 until each new task is trained, and obtaining the trained multi-layer convolutional neural network model with the conditional convolution structure;

[0016] S8: using the multi-layer convolutional neural network model with conditional convolution structure after training to perform target recognition on actual radar high-resolution range profile.

[0017] In an embodiment of the present application, the S1 comprises:

[0018] S1.1: obtaining a training sample set composed of a set of radar high-resolution range profile samples of multiple target aircraft models, each target in the training sample having a specific attitude angle;

[0019] S1.2: task division of the training samples in the training sample set according to the attitude angle range to obtain N tasks, each task containing all samples within a specific attitude angle range, and the attitude angle range between tasks has no repetition.

[0020] In an embodiment of the present application, the multi-layer convolutional neural network model with conditional convolution structure comprises three feature extractors and a fully connected classifier connected in sequence, wherein,

[0021] The first feature extractor comprises a conditional convolution layer, a pooling layer, an activation function layer and a normalization layer connected in sequence;

[0022] The second feature extractor comprises a conditional convolution layer, a pooling layer, an activation function layer and a normalization layer connected in sequence;

[0023] The third feature extractor comprises a conditional convolution layer, a pooling layer and an activation function layer connected in sequence;

[0024] The fully connected classifier is a fully connected classifier with correlation mapping mask.

[0025] In an embodiment of the present application, each conditional convolution layer comprises a plurality of expert convolution kernels and a fully connected layer, the fully connected layer being used to generate combination coefficients α0~α k of the input data through the fully connected layer, and the operation of the conditional convolution layer being:

[0026] CondConv(x)=σ((α0W0+α1W1+…+α k W k )*x)

[0027] Wherein, α=(α0α1…α k ) represents the adaptive routing parameters of the conditional convolution layer, i.e. the combination coefficients of the plurality of expert convolution kernels, W i represents the i-th expert convolution kernel, σ is the activation function, x is the input data, each expert convolution kernel corresponds to a divided task, and the parameter is updated only during training of the current task.

[0028] In one embodiment of the present application, the calculation rule of the full connection classifier with the relevance mapping mask is:

[0029] result=(W fcn ⊙soft_round(M i ,β)))·f

[0030] wherein f is a one-dimensional form of the features extracted by the feature extractor, W fcn represents the weight of the full connection classifier, soft_round(M i ,β) represents a soft binarization operation on the mask M i of the i-th task, and β is a constant adjusting the soft binarization strength.

[0031] In one embodiment of the present application, the S2 comprises:

[0032] S2.1: randomly set a training number for each of the divided N tasks, respectively 0-N-1;

[0033] S2.2: construct an initial multi-layer convolutional neural network model, train the initial multi-layer convolutional neural network model by the training sample set with the task number 0, and stop training after convergence is reached;

[0034] S2.3: convert the structure of the initial multi-layer convolutional neural network model, copy the convolution kernel parameters in the initial multi-layer convolutional neural network model to the expert convolution kernel of the initial task corresponding to the multi-layer convolutional neural network model with conditional convolution structure, and set the weight of the routing full connection network to 0 and the bias to 1;

[0035] S2.4: continue to train the multi-layer convolutional neural network model with conditional convolution structure by the training sample set with the task number 0, and set the part of the relevance mapping mask element corresponding to the current task of the classifier to zero after soft binarization processing during training until the number of training iterations is exhausted.

[0036] In one embodiment of the present application, the S3 comprises:

[0037] S3.1: expand the expert convolution kernel of each conditional convolution layer in the feature extractor of the multi-layer convolutional neural network model with conditional convolution structure, add an expert convolution kernel, and add a column of elements in the routing full connection layer, and the operation expression of the expanded conditional convolution layer is:

[0038] CondConv(x)=σ((α0W0+α1W1+…+α k W k +α k+1 Wk+1 )*x);

[0039] S3.2: extend the full connection classifier with the correlation mask, and add a correlation mask for the current task.

[0040] In an embodiment of the present application, the S4 comprises:

[0041] S4.1: calculate the contribution value of the expert convolution kernel corresponding to each task to the current conditional convolution operation through the routing module of the conditional convolution layer. i ;

[0042] S4.2: combine the expert convolution kernel corresponding to each task by weighting with the contribution value to obtain the conditional convolution kernel of the current task: i ;

[0043] W cond =α0W0+α1W+…+α k W k ;

[0044] S4.3: calculate the output result of the current conditional convolution layer through the conditional convolution kernel:

[0045] CondConv(x)=σ(W cond *x);

[0046] S4.4: initialize the correlation mask M corresponding to the current new task with a uniform distribution between 0 and 1. i ;

[0047] S4.5: soft binarize the correlation mask M through the S-shaped saturation function soft_round. i ;

[0048] S4.6: multiply the soft binarized mask soft_round(M i ) with the current model parameters to obtain the actually used model parameters, and calculate the actual output result of the model;

[0049] S4.7: obtain the gradients of the expert convolution kernel and the routing module parameters corresponding to the current task, the classifier weights, and the correlation mask of the current task by relying on the back propagation algorithm, and update the parameter values thereof.

[0050] S4.8: sum the binarized masks corresponding to the previously trained tasks to obtain the frozen parameter mask corresponding to the previous tasks:

[0051]

[0052] wherein binary(·) represents binarization, and Mt a mask representing the t-th task trained before;

[0053] S4.9: selecting the model parameters using the frozen parameter mask, and restoring them in the current model so as to keep them unchanged.

[0054] In an embodiment of the present application, the S5 comprises:

[0055] When the training accuracy of step S4 reaches the threshold value, the correlation mapping mask is pruned to sparsify the model parameters, and the elements in the soft binarized mask soft_round(M i ) that are less than the set threshold value are set to 0, so as to mask the elements corresponding to the current mask during model calculation, so that they do not participate in the calculation, wherein the binarization operation of the mask is:

[0056]

[0057] wherein m represents a single element of the mask, and threshold represents the set threshold value.

[0058] In an embodiment of the present application, the S8 comprises:

[0059] The target attitude angle relative to the radar is calculated by fitting the target track corresponding to the high-resolution range profile obtained by the radar;

[0060] According to the correspondence between the attitude angle and the task number, the task label to which the current to-be-identified data belongs is inferred;

[0061] The current to-be-identified data is input into the multi-layer convolutional neural network model with conditional convolution structure after training, and the model selects appropriate parameters to infer to obtain the recognition result.

[0062] Compared with the prior art, the present application has the following advantages:

[0063] 1. The conditional convolution layer is used to extract the features of HRRP in the present application, and the model capacity is improved by expanding the number of expert convolution kernels of the conditional convolution layer, but a convolution kernel is generated to adapt to the input data during calculation to perform convolution operation, which overcomes the problem that the capacity and calculation amount of the ordinary convolution layer increase linearly when the number of convolution kernels is expanded, so that the present application maintains the calculation amount similar to that of a single convolution kernel but the model capacity can be fully improved.

[0064] 2. The correlation mapping mask is used to cover the full connection layer classifier of the model in the present application, so that one classifier can serve multiple tasks at the same time, which overcomes the problem that the number of model classifiers increases proportionally with the number of tasks in the general conditional convolution method, and improves the parameter utilization rate of the full connection layer classifier.

[0065] 3、The task label is combined with radar HRRP target recognition, and a continuous training model without catastrophic forgetting is realized.

[0066] The application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a flowchart of an HRRP recognition model online updating method based on conditional convolution and correlation mapping provided by an embodiment of the application;

[0068] Figure 2 is another flowchart of an HRRP recognition model online updating method based on conditional convolution and correlation mapping provided by an embodiment of the application;

[0069] Figure 3 is a structural schematic diagram of an initial multi-layer convolutional neural network model provided by an embodiment of the application;

[0070] Figure 4 is a processing schematic diagram of a correlation mapping mask provided by an embodiment of the application;

[0071] Figure 5 is a data soft binarization schematic diagram provided by an embodiment of the application;

[0072] Figure 6 is a structural schematic diagram of a conditional convolution layer provided by an embodiment of the application;

[0073] Figure 7 is an expanded schematic diagram of a conditional convolution layer provided by an embodiment of the application. DETAILED DESCRIPTION

[0074] In order to further illustrate the technical means and effects adopted by the application to achieve the predetermined purposes, the HRRP recognition model online updating method based on conditional convolution and correlation mapping according to the application is described in detail below in combination with the accompanying drawings and specific embodiments.

[0075] The foregoing and other technical contents, features and effects of the application can be clearly presented in the detailed description of the specific embodiments below in combination with the accompanying drawings. Through the description of the specific embodiments, the technical means and effects adopted by the application to achieve the predetermined purposes can be understood more deeply and specifically. However, the accompanying drawings are provided for reference and illustration only, and are not used to limit the technical solutions of the application.

[0076] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a vesicle or an apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such vesicle or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the presence of additional identical elements in the vesicle or apparatus that includes the a recited element.

[0077] The embodiment of the present application provides a HRRP identification model online updating method based on conditional convolution and correlation mapping, please refer to Figure 1 and Figure 2 The HRRP identification model online updating method comprises the following steps.

[0078] S1: obtaining a training sample set with a target model label and task dividing the training sample set according to an attitude angle range of the target.

[0079] Specifically, a radar high-resolution range profile sample set of all known target aircraft models is obtained to form a training sample set, wherein each training sample has a specific attitude angle, and the attitude angle refers to an angle of rotation of a nose of the target aircraft clockwise to a radar line-of-sight direction. Then, the training samples in the training sample set are task divided according to the attitude angle range, wherein one task contains all samples in a range of attitude angles, the attitude angle range between tasks is not repeated, and the specific range size is determined according to the total attitude angle range and the sample quantity. In order to facilitate subsequent introduction, the total task quantity is denoted as N, the number of a single task is denoted as t, and the range of t is 0 to N-1.

[0080] S2: constructing an initial multi-layer convolutional neural network model and converting the initial multi-layer convolutional neural network model into a multi-layer convolutional neural network model with a conditional convolution structure through pre-training, the initial multi-layer convolutional neural network model comprising a feature extractor and a full-connection classifier with a mask structure.

[0081] The initial multi-layer convolutional neural network model comprises a feature extractor and a classifier, wherein the feature extractor is a convolutional neural network composed of ordinary convolutional layers, and the classifier is a full-connection classifier with a mask structure.

[0082] Specifically, please refer to Figure 3 , Figure 3is a structural schematic diagram of an initial multi-layer convolutional neural network model provided by an embodiment of the present application. The initial multi-layer convolutional neural network model of the present embodiment comprises three feature extractors and a fully connected classifier connected in sequence, wherein the first feature extractor comprises a conditional convolutional layer, a pooling layer, an activation function layer and a normalization layer connected in sequence; the second feature extractor comprises a conditional convolutional layer, a pooling layer, an activation function layer and a normalization layer connected in sequence; the third feature extractor comprises a conditional convolutional layer, a pooling layer and an activation function layer connected in sequence; and the fully connected classifier is a fully connected classifier with a mask structure.

[0083] In other words, the multi-layer convolutional neural network model with a conditional convolutional structure comprises a first feature extractor, a second feature extractor, a third feature extractor and a fully connected classifier, wherein the first feature extractor comprises a first convolutional layer, a first pooling layer, a first activation function layer and a first normalization layer connected in sequence; the second feature extractor comprises a second convolutional layer, a second pooling layer, a second activation function layer and a second normalization layer connected in sequence; the third feature extractor comprises a third convolutional layer, a third pooling layer and a third activation function layer connected in sequence; and the fully connected classifier is a fully connected classifier with a mask structure.

[0084] Further, the parameters of each module in the initial multi-layer convolutional neural network model of the present embodiment are shown in Table 1.

[0085] Table 1: Initial multi-layer convolutional neural network model

[0086]

[0087]

[0088] It should be noted that in other embodiments, the number of layers of the feature extractor in the initial multi-layer convolutional neural network model is not limited to three layers, but can also be other numbers.

[0089] The initial multi-layer convolutional neural network model will be converted into a multi-layer convolutional neural network model with a conditional convolutional structure after pre-training, and the difference between the multi-layer convolutional neural network model with a conditional convolutional structure and the initial multi-layer convolutional neural network model is that the convolutional layer of the model becomes a conditional convolutional layer.

[0090] In other words, the multi-layer convolutional neural network model with conditional convolution structure in the embodiment also comprises a first feature extractor, a second feature extractor, a third feature extractor and a full connection classifier connected in sequence, the first feature extractor comprises a first conditional convolution layer, a first pooling layer, a first activation function layer and a first normalization layer connected in sequence; the second feature extractor comprises a second conditional convolution layer, a second pooling layer, a second activation function layer and a second normalization layer connected in sequence; the third feature extractor comprises a third conditional convolution layer, a third pooling layer and a third activation function layer connected in sequence; and the full connection classifier is a full connection classifier with a correlation mapping mask. The specific structure can be referred to Table 2.

[0091] Table 2 multi-layer convolutional neural network model with conditional convolution structure

[0092]

[0093]

[0094] The conditional convolution layer comprises a plurality of expert convolution kernels and a full connection layer, wherein the full connection layer is used to generate combination coefficients α0~α k Therefore, the combination coefficients α are adaptive to the input data. Further, referring to Figure 6 , Figure 6 is a structural schematic diagram of a conditional convolution layer provided by the embodiment of the application. The structure of the conditional convolution layer can be introduced through its conditional convolution operation, and the calculation formula of a conventional convolution layer is as follows:

[0095] Conv(x)=σ(W*x)

[0096] wherein σ is an activation function, W is a weight parameter of the convolution layer, i.e. a convolution kernel, x is input data, and * represents a convolution operation; and the operation of the conditional convolution layer is represented as:

[0097] CondConv(x)=σ((α0W0+α1W1+...+α k W k )*x)

[0098] wherein α=(α0α1…α k ) represents adaptive routing parameters of the conditional convolution layer, i.e. combination coefficients of the plurality of expert convolution kernels, W i represents the i-th expert convolution kernel, each expert convolution kernel corresponds to a task and can only be updated during the training of the task, and participates in combination calculation only and does not participate in updating at other times.

[0099] The full connection classifier structure with a mask structure is also illustrated by formula, and the classification operation of a conventional full connection classifier on the features extracted by the feature extractor is as follows:

[0100] Classify(f)=W fcn ·f

[0101] where f is the one-dimensional form of the features extracted by the feature extractor, i.e. the flattening operation, and W fcn denotes the weights of the classifier, and · denotes the matrix multiplication, as shown in Figure 4 The classification operation of the fully connected classifier with the mask structure is as follows:

[0102] Classify masked (f)=(W fcn ⊙M)·f

[0103] where M denotes the binary mask, which is a parameter with the same size as the weights W fcn and whose elements are either 0 or 1, and denotes the dot product operation, 0 means that the element at the corresponding position of W fcn does not participate in the calculation, and 1 means that the element at the corresponding position participates in the calculation, accordingly, each task has a dedicated classifier binary mask, which is used to select the weights of the classifier, when the model has not been trained for the current task, the mask is initialized as a mask matrix uniformly distributed between 0 and 1, therefore the mask cannot be directly multiplied by the weights W fcn , but needs to go through a soft binary operation, as follows:

[0104]

[0105] where m denotes a single element of the mask, and β is a constant that adjusts the intensity of the soft binary operation, the function is a sigmoid saturation function, after the soft binary operation, the values of the mask will tend to 0 and 1, and a small amount of elements between 0 and 1 will be retained in the middle part, and the transition speed is extremely large, therefore it is called soft binary, as shown in the attached Figure 5 The operation will continue to be applied to the mask during training, in order to make the mask differentiable, so that its gradient can be calculated using the backpropagation algorithm, and then the mask value is updated, the mask is called the relevance mapping mask during the training process, which indicates the relevance of each position parameter to the current task, therefore, the calculation rule of the fully connected classifier containing the relevance mapping mask is as follows:

[0106] result=(W fcn ⊙soft_round(M i ,β)))·f

[0107] where M i denotes the mask of the i-th task.

[0108] Similarly, in other embodiments, the number of layers of the feature extractor in the multi-layer convolutional neural network model with conditional convolution structure is not limited to three layers, but can also be other numbers, and in addition, the expert convolution kernel of each layer can also be selected with other appropriate sizes.

[0109] In particular, step S2 of the embodiment includes:

[0110] S2.1: Randomly set a training number for the divided N tasks, respectively 0-N-1, and in actual application, the part with complete data in the gesture range should be trained as the current task in priority;

[0111] S2.2: Construct an initial multi-layer convolutional neural network model, train the initial multi-layer convolutional neural network model through the training sample set with task number 0, and stop training after convergence is reached;

[0112] S2.3: Convert the structure of the initial multi-layer convolutional neural network model, copy the convolution kernel parameters in the initial multi-layer convolutional neural network model to the expert convolution kernel of the initial task corresponding to the multi-layer conditional convolutional neural network model, and set the weight of the routing fully connected network to 0 and the bias to 1 at this time, that is, for the first task, the regular convolution kernel of the initial model acts as the expert convolution kernel of the task;

[0113] S2.4: Continue to train the multi-layer convolutional neural network model with conditional convolution structure through the training sample set with task number 0, and set the part with the result less than the threshold value in the correlation mapping mask element corresponding to the current task of the classifier after soft binarization processing to zero during training, until the number of training iterations is exhausted, wherein the threshold value is a manually set hyperparameter, which is usually 0.999999.

[0114] S3: Expand the expert convolution kernel and routing module correlation mapping mask of the multi-layer convolutional neural network model with conditional convolution structure for the next new task.

[0115] In the embodiment, S3 includes:

[0116] S3.1: Expand the expert convolution kernel of each conditional convolution layer in the feature extractor of the multi-layer convolutional neural network model with conditional convolution structure, and the conditional convolution calculation after expansion only increases one expert convolution kernel and one corresponding routing module compared with before expansion, as shown in Figure 7 The operation expression of the expanded conditional convolution layer is:

[0117] CondConv(x)=σ((α0W0+α1W1+...+α k W k +α k+1 W k+1 )*x)

[0118] The contribution value of the expert convolution kernel corresponding to each task to the current conditional convolution kernel is obtained by calculating the data through the routing module of the conditional convolution layer i where i is the task number, and the routing module is a fully connected layer. The data is first subjected to global average pooling and then subjected to the fully connected layer to obtain the routing weight, as follows:

[0119] α i = σ(G(x)R i +b i )

[0120] where G represents global average pooling, represents that the average value of the feature map of each channel is obtained, R i and b i represent the learnable routing weight and bias, respectively.

[0121] S3.2: Expand the fully connected classifier with the correlation mapping mask. The expansion operation only needs to add a correlation mapping mask for the current task.

[0122] S4: Train the expanded multi-layer convolutional neural network model with the conditional convolution structure for the current new task.

[0123] Specifically, the step S4 includes:

[0124] S4.1: Obtain the contribution value of the expert convolution kernel corresponding to each task to the current conditional convolution operation by calculating the input data through the routing module of the conditional convolution layer i , that is, the combination coefficient;

[0125] S4.2: Weighted combination of the expert convolution kernel corresponding to each task to obtain the conditional convolution kernel of the current task: i

[0126] W cond = α0W0+ α1W+ … + α k W k ;

[0127] S4.3: Calculate the output result of the data to the current convolution layer through the conditional convolution kernel, as follows:

[0128] CondConv(x) = σ(W cond *x)

[0129] S4.4: Initialize the correlation mapping mask M i corresponding to the current new task with a uniform distribution between 0 and 1.

[0130] ​S4.5: The correlation mapping mask M is processed by a sigmoid saturation function called soft_round, i.e., a soft rounding function i to perform soft binarization;

[0131] S4.6: The soft binarized mask soft_round(M i ) is multiplied by the current model parameters to obtain the actually used model parameters, and the actual output result is calculated;

[0132] S4.7: The expert convolution kernel corresponding to the current task, the routing module parameter, the classifier weight, and the gradient of the correlation mapping mask of the current task are obtained by relying on the back propagation algorithm, and the parameter values are updated;

[0133] It should be noted that the expert convolution kernel parameters of the non-current task are kept frozen, and only the expert convolution kernel parameters W i of the current task are updated.

[0134] S4.8: The binarized masks corresponding to the previously trained tasks are summed to obtain the frozen parameter mask corresponding to the previous tasks:

[0135]

[0136] where binary(·) represents soft binarization, M t represents the mask of the tth previously trained task.

[0137] S4.9: The model parameters are selected using the frozen parameter mask and restored to the current model, so that they remain unchanged.

[0138] It should be noted that the unchanged parameters are fixed parameters belonging to the previous old tasks, but participate in the calculation of the current task. The specific parameter selection is adaptively adjusted with the task correlation in the training process of the mask, which ensures the parameter reuse rate of similar tasks and the isolation degree of dissimilar task parameters.

[0139] S5: The correlation mapping mask of the full connection classifier is pruned using a model pruning method.

[0140] When the training accuracy of step S4 reaches a threshold, the correlation mapping mask is pruned to sparsify the model parameters, and the elements in the soft binarized mask soft_round(M i ) that are less than a set threshold are set to 0 to mask the elements corresponding to the current mask during model calculation, so that they do not participate in the calculation. The binarization operation of the mask is as follows:

[0141]

[0142] Wherein, m represents a single element of the mask, and threshold represents a set threshold.

[0143] S6: Perform performance recovery training on the pruned multi-layer convolutional neural network model using the current new task.

[0144] Specifically, continue to inject the data of the current new task into the model for training in the case where the mask has been pruned to restore certain performance, and determine whether the performance of the model on the current task meets the standard according to the validation dataset. If not, increase the recovery training time until the upper limit is reached.

[0145] S7: Obtain another new task and re-execute steps S3 to S6 until each new task is trained, and obtain the multi-layer convolutional neural network model with conditional convolutional structure after training.

[0146] Specifically, after each task training is completed, determine whether the recognition performance of the model meets the standard according to the validation set. If it meets the standard, execute S3 for the next new task. Otherwise, increase the recovery training time until the upper limit is reached. If there is no new task, execute S8.

[0147] S8: Use the multi-layer convolutional neural network model with conditional convolutional structure after training to perform target recognition on actual radar high-resolution range profiles.

[0148] Specifically, when the model is used online for recognition, the target track obtained by the radar is fitted to obtain the attitude angle of the target relative to the radar, and then the task label to which the current to-be-recognized data belongs is inferred according to the correspondence between the attitude angle and the task number. The current to-be-recognized data is input into the multi-layer convolutional neural network model with conditional convolutional structure after training. The model will select appropriate sparse parameters for inference to obtain the recognition result when reasoning. This step combines this method with radar HRRP target recognition, i.e., describes the availability of the task label in radar HRRP target recognition.

[0149] The method aims to continuously train the radar HRRP target recognition model under the target pose expansion scene, realize the learning and memory of different pose data tasks of the continuously incoming target, and mainly allocate different model parameters to different tasks and isolate the learnable parameters between different tasks to realize the learning and memory through the expansion network structure.

[0150] The application uses a conditional convolution layer to extract the features of the HRRP, expands the number of expert convolution kernels of the conditional convolution layer to improve the model capacity, but generates a convolution kernel for convolution operation according to the input data in the calculation process, overcomes the problem that the capacity and calculation amount of the ordinary convolution kernel increase linearly when the ordinary convolution kernel is expanded, makes the application maintain the calculation amount similar to that of a single convolution kernel, and the model capacity can be fully improved. The application uses a correlation mapping mask to cover the full connection layer classifier of the model, so that one classifier can serve multiple tasks at the same time, overcomes the problem that the number of model classifiers increases in proportion to the number of tasks in the general conditional convolution method, and improves the parameter utilization rate of the full connection layer classifier.

[0151] The application combines the task label and the radar HRRP target recognition organically, and realizes the continuously trained model without catastrophic forgetting.

[0152] In several embodiments provided by the application, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0153] In addition, each function module in each embodiment of the application can be integrated in a processing module, or each module can exist physically, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of hardware plus software function module.

[0154] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all should be deemed as falling within the protection scope of the present application.

Claims

1. A method for online updating of a HRRP recognition model based on conditional convolution and correlation mapping, characterized in that, The method comprises the following steps: S1: obtaining a training sample set with a target model label and task dividing the training sample set according to a target attitude angle range; S2: constructing an initial multi-layer convolutional neural network model and converting it into a multi-layer convolutional neural network model with a conditional convolutional structure through pre-training, wherein the initial multi-layer convolutional neural network model comprises a feature extractor and a full connection classifier with a mask structure; S3: expanding the expert convolutional kernel, the routing module and the correlation mapping mask of the multi-layer convolutional neural network model with the conditional convolutional structure for the next new task; S4: training the expanded multi-layer convolutional neural network model with the conditional convolutional structure for the current new task; S5: pruning the correlation mapping mask of the full connection classifier by using a model pruning method; S6: performing performance recovery training on the pruned multi-layer convolutional neural network model by using the current new task; S7: obtaining another new task and re-executing steps S3 to S6 until each new task is trained, and obtaining the trained multi-layer convolutional neural network model with the conditional convolutional structure; S8: performing target recognition on an actual radar high-resolution range image by using the trained multi-layer convolutional neural network model with the conditional convolutional structure; The multi-layer convolutional neural network model with the conditional convolutional structure comprises three feature extractors and a full connection classifier connected in sequence, wherein The first feature extractor comprises a conditional convolutional layer, a pooling layer, an activation function layer and a normalization layer connected in sequence; The second feature extractor comprises a conditional convolutional layer, a pooling layer, an activation function layer and a normalization layer connected in sequence; The third feature extractor comprises a conditional convolutional layer, a pooling layer and an activation function layer connected in sequence; The full connection classifier is a full connection classifier with a correlation mapping mask; Each conditional convolutional layer comprises a plurality of expert convolutional kernels and a full connection layer; S3.1: expanding the expert convolutional kernel of each conditional convolutional layer in the feature extractor of the multi-layer convolutional neural network model with the conditional convolutional structure, adding an expert convolutional kernel and a column of elements in the routing full connection layer, and the operation expression of the expanded conditional convolutional layer is: S3.2: expanding the full connection classifier with the correlation mapping mask, and adding a correlation mapping mask for the current task. ， wherein, ( … ) denotes an adaptive routing parameter of the conditional convolutional layer, denotes the i th expert convolutional kernel, σ is an activation function, is the input data; S1.1: obtaining a radar high-resolution range image sample set comprising a plurality of target aircraft models to form a training sample set, and each target in the training sample set has a specific attitude angle; 2. The HRRP recognition model based on conditional convolution and correlation mapping online updating method according to claim 1, characterized in that, S1.2: task dividing the training samples in the training sample set according to the attitude angle range to obtain N tasks, each task containing all samples within a specific attitude angle range, and the attitude angle ranges of the tasks are not repeated. The calculation rule of the full connection classifier with the correlation mapping mask is: S2.2: constructing an initial multi-layer convolutional neural network model, training the initial multi-layer convolutional neural network model by using the training sample set with a task number of 0, and stopping training when convergence is achieved.

3. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 2, characterized in that, The full connection layer is used to generate combination coefficients through the full connection layer based on the input data The operation of the conditional convolution layer is: wherein, ( … ) denotes the adaptive routing parameters of the conditional convolution layer, i.e. the combination coefficients of multiple expert convolution kernels, denotes the i expert convolution kernel, σ is an activation function, is the input data, each expert convolution kernel corresponds to one task divided, and the parameter is updated only when the current task is trained.

4. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 3, characterized in that, ​ wherein, is a one-dimensional form of the features extracted by the feature extractor, denotes the weights of the fully connected classifier, denotes a mask for the i-th task i performs a soft binarization operation, is a constant that regulates the intensity of the soft binarization.​ 5. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 4, characterized in that, ​ S2.1: for division N Each task is randomly assigned a training number, ranging from 0 to... N -1; ​ S2.3: converting the structure of the initial multi-layer convolutional neural network model, copying the convolution kernel parameters in the initial multi-layer convolutional neural network model to the expert convolution kernel of the initial task corresponding to the multi-layer convolutional neural network model with conditional convolution structure, setting the weights of the routing fully connected network to 0 and the bias to 1; S2.4: continue training the multi-layer convolutional neural network model with conditional convolution structure by the training sample set with task number 0, and set the part with the result less than the threshold value in the correlation mapping mask element corresponding to the current task of the classifier to 0 in the training until the training iteration number is exhausted.

6. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 5, characterized in that, The S4 comprises: S4.1: calculating the contribution value of the expert convolution kernel corresponding to each task to the current conditional convolution operation through the routing module of the conditional convolution layer on the input data ; S4.2: The expert convolution kernels corresponding to each task are processed... Weighted combination yields the conditional convolution kernel for the current task: ； S4.3: calculating the output result of the current conditional convolution layer by the conditional convolution kernel: ； S4.4: initialize the relevance map mask corresponding to the current new task with a uniform distribution between 0 and 1 ; S4.5: by a sigmoid function correlation map mask soft binarization; S4.6: multiply the soft-binarized mask with the current model parameters to obtain the actually used model parameters, and calculate the actual output result of the model; S4.7: obtaining the gradient of the expert convolution kernel and the routing module parameter corresponding to the current task, the classifier weight, and the current task correlation mapping mask by the back propagation algorithm and updating the parameter value thereof; S4.8: summing the binarization mask corresponding to the previously trained task to obtain the frozen parameter mask corresponding to the previous task: ； wherein, denotes binarization, denotes a mask of the previously trained task; t denotes a mask of the previously trained task; S4.9: selecting the model parameter by using the frozen parameter mask and restoring it to the current model so as to keep it unchanged.

7. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 6, characterized in that, The S5 comprises: When the training accuracy of step S4 reaches the threshold value, the correlation mapping mask is pruned to sparsify the model parameters, and the soft binarized mask The elements less than the set threshold value are set to 0, so that the elements corresponding to the current mask are shielded during model calculation, and do not participate in the calculation, wherein the binarization operation of the mask is: ， wherein, m represents a single element of the mask, represents a set threshold value.

8. The HRRP recognition model online updating method based on conditional convolution and correlation mapping according to claim 7, characterized in that, The S8 comprises: calculating the attitude angle of the target relative to the radar by fitting the target track corresponding to the high-resolution range profile obtained by the radar; inferred the task label to which the current to-be-identified data belongs according to the correspondence between the attitude angle and the task number; inputting the current to-be-identified data into the multi-layer convolutional neural network model with conditional convolution structure trained, and the model selecting the appropriate parameter to infer the recognition result.

Images