Radar HRRP target recognition method and device based on attitude angle auxiliary information and electronic equipment

Abstract

The invention provides a radar HRRP target recognition method and device based on attitude angle auxiliary information and electronic equipment, relates to the technical field of radars, and aims to solve the problem that attitude angle tags of sample targets of a training set are available but attitude angle tags of a test set are missing in practical application. The method comprises the following steps: inputting radar HRRP data into a parameter sharing network of a target identification model to obtain feature information; inputting the feature information into a first specific task network to obtain an identification result of the target category; the target recognition model is obtained by training based on sample radar HRRP data, category labels and attitude angle labels of sample targets. During training, the two specific task networks jointly adopt the output information of the parameter sharing network, the output information is obtained based on the input sample radar HRRP data, the attitude angle of the to-be-recognized target does not need to be input, the attitude angle does not need to be input in the test stage, and the performance of the target recognition model is improved.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a radar HRRP target recognition method, device and electronic equipment based on attitude angle auxiliary information. Background Art

[0002] High Resolution Range Profile (HRRP) is a common tool for radar target recognition because of its low dimensionality, ease of acquisition and processing. HRRP reflects the position distribution of the target's radar cross section (RCS) in the direction of the radar's line of sight, and contains the structure and position information between each scatterer. The target's HRRP data can be used to train an end-to-end deep network to achieve feature extraction and category prediction of the target to be identified, and then execute a response plan based on the category prediction.

[0003] In the related art, the posture angle corresponding to the sample target is usually used as auxiliary information, and both the sample HRRP data and the posture angle corresponding to the sample target are used as input data of the target recognition model for training.

[0004] However, in the above-mentioned related technologies, when training the target recognition model, since both inputs need to be used as input data, both the training set and the test set need to have attitude angles. However, the echo received by the radar in the real scene does not contain the corresponding attitude angle. Therefore, in the test phase, the input corresponding to the attitude angle will be left blank, which will result in the missing attitude angle information in the test phase, thereby reducing the performance of the target recognition model finally trained. Summary of the invention

[0005] The present invention provides a radar HRRP target recognition method, device and electronic equipment based on attitude angle auxiliary information, which are used to solve the defect in the prior art that the performance of the target recognition model finally obtained by training is reduced.

[0006] The present invention provides a radar HRRP target recognition method based on attitude angle auxiliary information, comprising the following steps.

[0007] Obtain radar HRRP data of the target to be identified; Inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; Inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0008] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the target recognition model is trained based on the following method: Acquiring sample radar HRRP data of the sample target; Inputting the sample radar HRRP data into the parameter sharing network of the initial target recognition model to obtain sample feature information output by the parameter sharing network; Inputting the sample feature information into the first specific task network and the second specific task network of the initial target recognition model respectively, to obtain the predicted category output by the first specific task network and the predicted posture angle output by the second specific task network; Determining a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target; Based on the total loss function, the shared parameters of the parameter sharing network, the network parameters of the first specific task network, and the network parameters of the second specific task network are iteratively adjusted to obtain the target recognition model.

[0009] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the total loss function is determined based on the predicted category, the category label of the sample target, the predicted attitude angle and the attitude angle label corresponding to the sample target, including: Determining a recognition task loss function based on the predicted category and the category label of the sample target; Determining an angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target; The total loss function is determined based on the following formula (1): (1) in, represents the total loss function, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the angle estimation task loss function, Represents the recognition task loss function.

[0010] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the obtaining of sample radar HRRP data of the sample target includes: The optimization problem is constructed based on the following formula (2): (2) The constraints are , Represents the total number of sample targets, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the sample selection threshold, Indicates The control parameters of the sample target, It is expressed by the following formula (3): (3) represents the regularization term, , Indicates Multi-task joint loss of sample targets, , represents the weight of the angle estimation task, Indicates The angle estimation task loss function corresponding to the sample target is: Indicates The pose angle labels corresponding to the sample targets are Indicates The predicted posture angle corresponding to the sample target, represents the network parameters of the second task-specific network, Indicates The recognition task loss function corresponding to the sample target is: Indicates Sample targets correspond to classification labels, Indicates The predicted category corresponding to the sample target, representing network parameters of a first task-specific network; Solve the optimization problem to obtain the network parameters of the target recognition model .

[0011] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the angle estimation task loss function is determined based on the predicted attitude angle and the attitude angle label corresponding to the sample target, including: The interval width is determined based on the following formula (4): (4) in, represents the width of the interval, Indicates the maximum angle within the angle space, Indicates the minimum angle within the angle space range, Indicates the total number of pre-set intervals, represents a mapping function; The discrete angle corresponding to the predicted posture angle of the sample target is determined based on the following formula (5): (5) in, represents the predicted attitude angle The corresponding discrete angle; The distance-penalized metric of the discrete angle is determined based on the following formula (6): (6) in, represents the metric after distance penalty, represents the serial number of the interval in which the predicted attitude angle is located, represents weight; The discrete angle is determined based on the following formula (7): Distance penalty error relative to different intervals: (7) in, Represents the discrete angle Relative to the distance penalty error of different intervals, Indicates the sequence number of the kth interval; The angle estimation task loss function is determined based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target.

[0012] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the angle estimation task loss function is determined based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target, including: Mapping the distance penalty error of the discrete angle to obtain the probability density of the discrete angle in different intervals; The angle estimation task loss function is determined based on the following formula (8): (8) in, represents the angle estimation task loss function, Represents the total number of sample targets, Indicates The probability density of the discrete angles of sample targets in different intervals, Indicates The vector of pose angle labels of sample targets, Indicates The cosine weight of the sample target, , Indicates The pose angle labels of the sample targets.

[0013] According to a radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention, the parameter sharing network includes a first convolution module, a second convolution module, a third convolution module and a first fully connected layer connected in sequence, the first convolution module, the second convolution module and the third convolution module all include a convolution layer, a batch normalization layer, a first activation function layer and a maximum pooling layer connected in sequence, and the first specific task network and the second specific task network all include a second fully connected layer, a second activation function layer and a third fully connected layer connected in sequence.

[0014] The present invention also provides a radar HRRP target recognition device based on attitude angle auxiliary information, comprising: An acquisition unit, used for acquiring radar HRRP data of a target to be identified; An extraction unit, used for inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; A recognition unit, used for inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the radar HRRP target recognition method based on attitude angle auxiliary information as described in any one of the above is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the radar HRRP target recognition method based on attitude angle auxiliary information as described in any one of the above is implemented.

[0017] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for radar HRRP target recognition based on attitude angle auxiliary information as described above is implemented.

[0018] The radar HRRP target recognition method, device and electronic device based on attitude angle auxiliary information provided by the present invention input the acquired radar HRRP data of the target to be identified into the parameter sharing network of the target recognition model to obtain the feature information output by the parameter sharing network, and then input the feature information into the first specific task network of the target recognition model to obtain the recognition result of the target category output by the first specific task network. Among them, the target recognition model is obtained by training the initial target recognition model based on the sample radar HRRP data of the sample target, the category label of the sample target and the attitude angle label corresponding to the sample target. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle. It can be seen that when training the target recognition model of the present invention, the first specific task network and the second specific task network jointly use the output information of the parameter sharing network, and the output information is obtained based on the input sample radar HRRP data, and does not require the input of the attitude angle of the target to be identified. The attitude angle of the target to be identified is only used as a label, so there is no need to input the attitude angle in the test phase, thereby avoiding the problem of reduced performance of the target recognition model due to the lack of attitude angle information in the test phase in the related technology, thereby improving the performance of the target recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a schematic diagram of a HRRP neural network training method based on auxiliary information in the related art.

[0021] Figure 2 It is a flow chart of a radar HRRP target recognition method based on attitude angle auxiliary information provided by an embodiment of the present invention.

[0022] Figure 3 It is a schematic diagram of the training process of the target recognition model provided by an embodiment of the present invention.

[0023] Figure 4 Schematic diagram of the network structure of the initial target recognition model provided by the embodiment of the present invention.

[0024] Figure 5 It is a structural schematic diagram of a radar HRRP target recognition device based on attitude angle auxiliary information provided by an embodiment of the present invention.

[0025] Figure 6 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Figure 1 is a schematic diagram of the HRRP neural network training method based on auxiliary information in the related art, such as Figure 1 As shown, the HRRP neural network includes a convolution module 1, a convolution module 2 and a specific task layer connected in sequence, wherein the convolution module 1 and the convolution module 2 both include a convolution layer, a batch normalization layer, an activation function layer 1 and a maximum pooling layer connected in sequence, the specific task layer includes a fully connected layer 1, an activation function layer 2 and a fully connected layer 2 connected in sequence, the input of the convolution module 1 includes the sample HRRP data of the sample target and the attitude angle corresponding to the sample HRRP data, and the output of the specific task layer is the category of the sample target. However, in the related art, when training the HRRP neural network, since the sample HRRP data and the attitude angle corresponding to the sample HRRP data are both required as input data, the training set and the test set are required to have attitude angles, but the echo received by the radar in the real scene does not contain the corresponding attitude angle. Therefore, in the test phase, the input corresponding to the attitude angle will be left blank, which will result in the lack of attitude angle information in the test phase, thereby resulting in a reduction in the performance of the target recognition model finally trained.

[0028] Based on this, the present invention proposes a radar HRRP target recognition method based on attitude angle auxiliary information, inputting the acquired radar HRRP data of the target to be identified into the parameter sharing network of the target recognition model, obtaining the feature information output by the parameter sharing network, and then inputting the feature information into the first specific task network of the target recognition model to obtain the recognition result of the target category output by the first specific task network. Among them, the target recognition model is obtained by training the initial target recognition model based on the sample radar HRRP data of the sample target, the category label of the sample target and the attitude angle label corresponding to the sample target. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle. It can be seen that when training the target recognition model of the present invention, the first specific task network and the second specific task network jointly adopt the output information of the parameter sharing network, and the output information is obtained based on the input sample radar HRRP data, and does not require the input of the attitude angle of the target to be identified. The attitude angle of the target to be identified is only used as a label. Therefore, there is no need to input the attitude angle of the target to be identified in the test phase, thereby avoiding the problem of reduced performance of the target recognition model caused by the lack of attitude angle information in the test phase in the related technology, thereby improving the performance of the target recognition model.

[0029] Combine the following Figure 2-Figure 4 The present invention describes a radar HRRP target recognition method based on attitude angle auxiliary information. The execution subject of the radar HRRP target recognition method based on attitude angle auxiliary information can be an electronic device such as a terminal, a tablet computer or a computer, or a radar HRRP target recognition device based on attitude angle auxiliary information set in the electronic device. The radar HRRP target recognition device based on attitude angle auxiliary information can be implemented by software, hardware or a combination of the two.

[0030] Figure 2 FIG. 1 is a flow chart of a radar HRRP target recognition method based on attitude angle auxiliary information provided by an embodiment of the present invention. Figure 2 As shown, the radar HRRP target recognition method based on attitude angle auxiliary information includes the following steps: Step 201: Acquire radar HRRP data of a target to be identified.

[0031] For example, the acquisition and processing of HRRP data is of great significance for target recognition, because HRRP data contains information related to the size of the target, the physical structure of the target, and the distribution of scattering points. In practical applications, HRRP data can be acquired by microwave darkroom measurement.

[0032] Step 202: input the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network.

[0033] Among them, the parameter sharing network includes a first convolution module, a second convolution module, a third convolution module and a first fully connected layer connected in sequence, and the first convolution module, the second convolution module and the third convolution module each include a convolution layer, a batch normalization layer, a first activation function layer and a maximum pooling layer connected in sequence.

[0034] For example, the radar HRRP data is input into the parameter sharing network, and feature extraction is performed through the first convolution module, the second convolution module, and the third convolution module in sequence. The output of the first convolution module can be expressed by the following formula (9):

[0035] in, represents the output of the first convolutional module, Represents radar HRRP data, represents the output of the convolutional layer of the first convolutional module, represents the output of the batch normalization layer of the first convolutional module, Represents the activation function.

[0036] After obtaining the output of the third convolution module, the output of the third convolution module is subjected to feature dimension reduction through the first fully connected layer with dropout, and finally the feature information output by the first fully connected layer is obtained, that is, the feature information output by the parameter sharing network. The first convolution module, the second convolution module, the third convolution module and the first fully connected layer together constitute a parameter sharing network. The shared parameters of the parameter sharing network are .

[0037] Step 203: input the feature information into the first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network.

[0038] Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0039] For example, when the feature information output by the parameter sharing network is obtained, the feature information is input into the first specific task network of the target recognition model, and the recognition result of the target category is output through the first specific task network. Of course, the feature information can also be input into the second specific task network of the target recognition model, and the attitude angle estimation of the target to be identified is output through the second specific task network. If attitude angle estimation and category recognition are regarded as two tasks, the first task is t The network parameters of a task-specific network are , then t The output of a task-specific network It can be expressed by the following formula (10): (10) in, Indicates t The network parameters of a task-specific network, Represents the characteristic information output by the parameter sharing network, Indicates t The bias term of the last fully connected layer in a task-specific network.

[0040] It should be noted that the present invention does not limit the number of convolution modules included in the parameter sharing network. For example, there may be four convolution modules, etc., which can be set based on demand.

[0041] The radar HRRP target recognition method based on attitude angle auxiliary information provided by the present invention inputs the acquired radar HRRP data of the target to be identified into the parameter sharing network of the target recognition model to obtain the feature information output by the parameter sharing network, and then inputs the feature information into the first specific task network of the target recognition model to obtain the recognition result of the target category output by the first specific task network. Wherein, the target recognition model is obtained by training the initial target recognition model based on the sample radar HRRP data of the sample target, the category label of the sample target and the attitude angle label corresponding to the sample target, and the initial target recognition model includes the parameter sharing network, the first specific task network and the second specific task network, both of which are connected to the parameter sharing network, the first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle. It can be seen that when training the target recognition model of the present invention, the first specific task network and the second specific task network jointly use the output information of the parameter sharing network, and the output information is obtained based on the input sample radar HRRP data, and does not require the input of the attitude angle of the target to be identified. Only the attitude angle of the target to be identified is used as a label as auxiliary information. Therefore, there is no need to input the attitude angle of the target to be identified in the test phase, thereby avoiding the problem of reduced performance of the target recognition model caused by the lack of attitude angle information in the test phase in the related technology, thereby improving the performance of the target recognition model.

[0042] In one embodiment, Figure 3 is a schematic diagram of the training process of the target recognition model provided by the embodiment of the present invention, such as Figure 3 As shown, the target recognition model is trained based on the following method: Step 301: Acquire sample radar HRRP data of the sample target.

[0043] For example, the sample radar HRRP data of each sample target may be acquired by a microwave darkroom measurement method.

[0044] Step 302: Input the sample radar HRRP data into the parameter sharing network of the initial target recognition model to obtain sample feature information output by the parameter sharing network.

[0045] For example, when the sample radar HRRP data of each sample target is obtained, each sample radar HRRP data is input into the parameter sharing network of the initial target recognition model, and feature extraction is implemented in sequence through the first convolution module, the second convolution module and the third convolution module of the initial target recognition model, and feature dimension reduction is performed through the first fully connected layer with dropout, and finally the sample feature information output by the parameter sharing network of the initial target recognition model is obtained.

[0046] Step 303: input the sample feature information into the first specific task network and the second specific task network of the initial target recognition model respectively, to obtain the predicted category output by the first specific task network and the predicted posture angle output by the second specific task network.

[0047] For example, when the sample feature information output by the parameter sharing network of the initial target recognition model is obtained, the sample feature information is input into the first specific task network of the initial target recognition model, the predicted category is output through the first specific task network, and the sample feature information is input into the second specific task network of the initial target recognition model, and the predicted attitude angle is output through the second specific task network.

[0048] Step 304: Determine a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target.

[0049] For example, when the predicted category and predicted posture angle of the sample target are obtained, a total loss function is constructed based on the difference between the predicted category of the sample target and the category label of the sample target, and the difference between the predicted posture angle of the sample target and the posture angle label corresponding to the sample target.

[0050] Step 305: Based on the total loss function, iteratively adjust the shared parameters of the parameter sharing network, the network parameters of the first specific task network, and the network parameters of the second specific task network to obtain the target recognition model.

[0051] For example, when the total loss function is obtained, the shared parameters of the parameter sharing network, the network parameters of the first specific task network, and the network parameters of the second specific task network can be iteratively adjusted based on the total loss function until the convergence conditions are reached, and finally the target recognition model is obtained.

[0052] It should be noted that in each iteration, the gradient of the shared parameter is the weighted sum of the gradients of multiple tasks, which can be specifically expressed by the following formula (11): (11) in, Represents shared parameters For the total loss function The gradient of Indicates the number of tasks, here , Indicates The weight of the task, Indicates The loss function for each task.

[0053] Shared Parameters Updated by the gradient descent process: ,in, Represents the learning rate. In the training phase, it is necessary to calculate the recognition task loss function based on the predicted category and category label, and also to calculate the angle estimation task loss function based on the predicted attitude angle and attitude angle label. By weighting the recognition task loss function and the angle estimation task loss function, the multi-task joint loss, that is, the total loss function, can be obtained. In the prediction phase, only the radar HRRP data of the target to be identified needs to be input to obtain the recognition result of the target category and the attitude angle estimation of the target to be identified.

[0054] Figure 4 is a schematic diagram of the network structure of the initial target recognition model provided by an embodiment of the present invention, such as Figure 4 As shown, the initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network, wherein the first specific task network and the second specific task network are both connected to the parameter sharing network, the parameter sharing network includes a first convolution module, a second convolution module, a third convolution module and a first fully connected layer connected in sequence, the first convolution module, the second convolution module and the third convolution module all include a convolution layer, a batch normalization layer, a first activation function layer and a maximum pooling layer connected in sequence, the first specific task network and the second specific task network all include a second fully connected layer, a second activation function layer and a third fully connected layer connected in sequence, the input of the first convolution module is the sample radar HRRP data, the first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle, based on the predicted category, the category label of the sample target, the predicted attitude angle and the attitude angle label corresponding to the sample target, the total loss function is determined, based on the total loss function, the shared parameters of the parameter sharing network, the network parameters of the first specific task network and the network parameters of the second specific task network are iteratively adjusted to obtain the target recognition model. The parameter sharing network adopts a hard parameter sharing architecture. Since the recognition task and the angle estimation task are highly correlated, the use of a hard parameter sharing architecture can reduce the amount of parameters and computing resources without reducing the feature extraction effect.

[0055] It should be noted that the initial target recognition model may also be referred to as an adaptive auxiliary learning network (Adaptive Auxiliary Learning Net, AALNet), which is not limited in the present invention.

[0056] In this embodiment, when training the target recognition model, the first specific task network and the second specific task network jointly use the output information of the parameter sharing network, and the output information is obtained based on the input sample radar HRRP data, and does not require the input of the attitude angle. Only the attitude angle, which is an auxiliary information, is used as a label. Therefore, there is no need to input the attitude angle in the test phase, thereby avoiding the problem of the lack of attitude angle information in the test phase in the related technology leading to the reduction of the performance of the target recognition model, and improving the performance of the trained target recognition model.

[0057] In one embodiment, the above step 304 determines the total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target, which can be specifically implemented in the following manner: Based on the predicted category and the category label of the sample target, determine the recognition task loss function; based on the predicted attitude angle and the attitude angle label corresponding to the sample target, determine the angle estimation task loss function; based on the following formula (1), determine the total loss function: (1) in, represents the total loss function, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the angle estimation task loss function, Represents the recognition task loss function.

[0058] For example, the present invention involves two tasks, namely, recognition task and angle estimation task. The recognition task is the main task and can be regarded as a classification problem. The angle estimation task is an auxiliary task and can be regarded as a regression problem. The joint objective function under multi-task conditions is derived based on Gaussian maximum likelihood with homoscedastic uncertainty. It is assumed that the initial target recognition model is input as sample radar HRRP data. , the network parameters are Under the condition of , two tasks have the same input, in the case of multiple task-specific network outputs, assuming that the task scenario satisfies two conditions: is a sufficient statistic, and Tags for tasks The conditional independence assumption is satisfied, so the probability density function of the multi-task output can be expressed as follows: (12) in, express The probability density function of the task output is represents the probability density function of the first task output, Indicates The probability density function of the task output.

[0059] Different types of tasks usually correspond to different output forms, so Will correspond to different distribution models. In maximum likelihood inference, it is necessary to maximize the likelihood function of the model and derive the weighted loss of multiple tasks when considering homoscedastic uncertainty.

[0060] For a regression problem, the output of the regression problem can be modeled as a Gaussian distribution with observation noise, which can be specifically expressed by the following formula (13): (13) in, Represents the uncertainty parameter, which is also a trainable parameter. As the scale of observation noise, it is fixed in the weight decay of the neural network, so it can be adjusted by the maximum likelihood inference method. , the maximum likelihood of the objective function is to minimize the negative log-likelihood of the model. Specifically, the above formula (13) can be solved by the following formula (14): (14) in, Indicates equivalence, Indicates the label corresponding to the task.

[0061] In addition, usually, the likelihood of the output is adjusted by using the softmax function. Therefore, for a classification problem, the output of the classification problem can be modeled as a Gibbs distribution with a temperature coefficient, which can be specifically expressed by the following formula (15): (15) Among them, the parameters Used to describe the flatness of a distribution. Often called the temperature coefficient.

[0062] The above formula (15) can be solved by the following formula (16): (16) in, represents the category label, Represents a collection of categories Any category in , yes An element in a vector.

[0063] In a multi-task scenario with 2 tasks, the two tasks are the recognition task of class prediction of sample targets and the angle estimation task of posture angle estimation of sample targets. The recognition task is regarded as a classification problem with discrete output, while the angle estimation task is regarded as a regression problem with continuous output. Therefore, continuous output needs to be obtained at the same time. and discrete outputs The joint loss function can be expressed as follows: (17) in, represents the joint loss function, that is, the total loss function, represents the network parameters of the target recognition model, and Represents a trainable parameter.

[0064] Furthermore, it is known that two independent tasks are expressed as The mean square error loss and The cross entropy loss At the same time, the following formula (18) can be assumed based on experience:

[0065] when When it approaches 1, the two are equal. Therefore, the total loss function can be expressed by the following formula (19): (19) in, represents the angle estimation task loss function, Represents the recognition task loss function, which dynamically adjusts the importance of the two tasks during network training.

[0066] Extending formula (19) to the general solution in a multi-task scenario where M outputs are modeled as Gaussian distribution and Gibbs distribution, it can be specifically expressed by the following formula (20): (20) Among them, K represents the number of regression tasks modeled as Gaussian distribution in the total number of M tasks, and the remaining tasks are all classification tasks modeled as Gibbs distribution, and the number is MK. Since the loss shown in the above formula (20) will cause the loss function to degenerate to 0, modifying the regularization term can meet the condition that the multi-task joint loss is non-negative, so the above formula (20) is further expressed as the following formula (21): (twenty one) Furthermore, since the present invention only includes two tasks, the above formula (21) can be expressed by the above formula (1). By expressing the total loss function of formula (1), it is easy to know that only the setting By initializing the parameters, it is possible to adaptively and dynamically adjust the weights of different tasks during training based on the amount of information contained in different tasks and the homoscedastic uncertainty, without the need to manually select the weights of the tasks.

[0067] In this embodiment, a loss function weighting method is designed when multiple tasks have a primary and secondary relationship, and the same variance uncertainty is used to evaluate the importance of different tasks, so as to better identify the primary and secondary relationships of multiple tasks and further improve the performance of the trained target recognition model.

[0068] In one embodiment, the above step 301 obtains the sample radar HRRP data of the sample target, which can be specifically implemented in the following manner: The optimization problem is constructed based on the following formula (2): (2) The constraints are , Represents the total number of sample targets, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the sample selection threshold, Indicates The control parameters of the sample target, It is expressed by the following formula (3): (3) represents the regularization term, , Indicates Multi-task joint loss of sample targets, , represents the weight of the angle estimation task, Indicates The angle estimation task loss function corresponding to the sample target is: Indicates The pose angle labels corresponding to the sample targets are Indicates The predicted posture angle corresponding to the sample target, represents the network parameters of the second task-specific network, Indicates The recognition task loss function corresponding to the sample target is: Indicates Sample targets correspond to classification labels, Indicates The predicted category corresponding to the sample target, Represents network parameters of the first task-specific network.

[0069] Solve the optimization problem to obtain the network parameters of the target recognition model .

[0070] For example, the recognition task and the angle estimation task of the present invention appear as the main task and the auxiliary task respectively. The priority of the tasks leads to the loss function being not reasonable according to the simple multi-task adaptive weighting. At the same time, in the early training of the network, if the weight of the auxiliary task is too large, it will cause the network training direction to deviate, making it easier to fall into the local optimal solution. Therefore, the present invention uses the Self-paced strategy to improve the weighted loss function that uses uncertainty as the task weight, that is, to improve the total loss function shown in the above formula (1), and at the same time proposes a sample selection method that meets this scenario. The self-paced strategy of the model is used to realize the selection order of training samples "easy first and difficult later". In each iteration, simple training samples are selected first and parameters are learned, which can effectively avoid the algorithm from converging to the local optimal state, reduce the difficulty of model training, and speed up the fitting speed.

[0071] In the application scenario of the present invention, the simple samples and difficult samples are defined as follows: when the category classification of the sample target is correct, and the difference between the predicted attitude angle and the attitude angle label does not exceed the preset threshold, the sample radar HRRP data of the sample target is defined as a simple sample; when the category classification of the sample target is wrong, regardless of the effect of the predicted attitude angle, the sample radar HRRP data of the sample target is defined as a difficult sample. The training set is defined as ,in, Indicates training samples, Indicates The true labels corresponding to the training samples are Indicates that the true label is When the input data The loss function is, Indicates The predicted output corresponding to the training samples is is the network parameter. Then we can get the optimization problem shown in the above formula (2). Since the main task contains more information, it should have a larger weight than the auxiliary task. Therefore, we introduce represents the weight of the auxiliary task, that is, the weight of the angle estimation task, for example, Set to 0.3, for the optimization problem of formula (2), the alternative convex search (ACS) can be used to implement it. Through the iterative double convex optimization method, , solve the optimization problem of formula (2) to obtain the network parameters of the target recognition model , when the network parameters are fixed When The global optimal solution of is expressed by the above formula (3). By judging whether the recognition task loss function of the main task is greater than the sample selection threshold, it is judged whether the sample radar HRRP data of the sample target is a difficult sample that is difficult to correctly identify. If the sample radar HRRP data is a difficult sample, then Take 0, that is, temporarily do not include the sample in the network training process. As the network training progresses, the threshold is set to increase with the training round cycle. It keeps increasing, so as to select the training samples from easy to difficult. In the early stage of training, it focuses on the selection of simple samples, so that the target recognition model is trained in a direction that is beneficial to the main task. In the middle and late stages of training, difficult samples are gradually introduced to strengthen the fine-tuning of the target recognition model by auxiliary tasks and improve the training results.

[0072] In this embodiment, a sample selection method is designed based on a self-paced strategy, so that the target recognition model is more focused on optimizing the main task during training, avoiding falling into a local optimum, and further improving the performance of the trained target recognition model.

[0073] In one embodiment, the above-mentioned determination of the angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target can be specifically implemented in the following manner: The interval width is determined based on the following formula (4): (4) in, represents the width of the interval, Indicates the maximum angle within the angle space, Indicates the minimum angle within the angle space range, Indicates the total number of pre-set intervals, represents a mapping function; The discrete angle corresponding to the predicted posture angle of the sample target is determined based on the following formula (5): (5) in, represents the predicted attitude angle The corresponding discrete angle; The distance-penalized metric of the discrete angle is determined based on the following formula (6): (6) in, represents the distance penalty metric, represents the serial number of the interval in which the predicted attitude angle is located, represents weight; The discrete angle is determined based on the following formula (7): Distance penalty error relative to different intervals: (7) in, Represents the discrete angle Relative to the distance penalty error of different intervals, Indicates The serial number of the interval; The angle estimation task loss function is determined based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target.

[0074] For example, the regression problem is a non-convex optimization problem, which makes it difficult for the model to obtain an accurate estimate of the unknown variables when the solution space is infinitely continuous, resulting in sub-optimization. At the same time, when estimating variables with periodic or ordinal properties such as attitude angles and discrete heights, the mean square error loss cannot well represent the gap between the predicted variables and the actual variables. In order to solve this problem, ordered regression is usually introduced to replace the regression problem to model the task. The present invention proposes an orientation prediction loss function based on cosine-weighted ordered regression, which replaces the original mean square error loss as the loss function of the auxiliary task in multi-task learning. Since it is difficult for the regression problem to accurately predict the target variable under the condition of infinite solution space, the present invention considers converting the infinite space of the solution of the regression problem into a finite discrete space of estimated probability density from the perspective of quantization. In essence, the regression problem is converted into a special classification problem by means of discretization. It is generally divided into the following five steps: identity mapping, interval partitioning, distance penalty measurement, probability distribution and cosine weighting.

[0075] First, determine that the angle space range to be predicted is 0 degrees to 360 degrees, then the maximum angle within the angle space range is 360 degrees, the minimum angle within the angular space is 0 degrees, and the angle is mapped to a discrete angle by an identity mapping method. Specifically, the interval width can be calculated by the above formula (4), and the interval width is substituted into the above formula (5) to calculate the discrete angle corresponding to the predicted posture angle of the sample target. Here The value of can be 36. Assume that the predicted attitude angle of the sample target is is 121 degrees, is 10 degrees, then =12.1.

[0076] Then, after getting When , the angle space range can be divided into intervals, which can be expressed as ,in, Indicates the first interval, the value is 0, indicating the sequence number of the first interval, Indicates interval, the value is , indicating the The serial number of the interval.

[0077] Then, the distance penalty metric is used to measure the discrete angle and Because the farther the true angle value is from the discrete angle interval, the greater the error, so the distance penalty of the discrete angle can be expressed by the above formula (6), and the discrete angle can be further obtained based on the above formula (7). Relative to the distance penalty error in different intervals, the angle estimation task loss function is determined based on the distance penalty error of discrete angles, the predicted attitude angle and the attitude angle label corresponding to the sample target.

[0078] In this embodiment, the predicted attitude angle of the sample target is discretized to obtain the discrete angle corresponding to the predicted attitude angle of the sample target, and the distance-penalized metric of the discrete angle is calculated. Then, based on the distance-penalized metric of the discrete angle, the distance penalty error of the discrete angle relative to different intervals is calculated. Finally, based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target, the angle estimation task loss function is determined to solve the problem of large estimation error caused by the infinite continuous solution space of the traditional regression problem.

[0079] In one embodiment, the above-mentioned determination of the angle estimation task loss function based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target can be specifically implemented in the following manner: Mapping the distance penalty error of the discrete angle to obtain the probability density of the discrete angle in different intervals; The angle estimation task loss function is determined based on the following formula (8): (8) in, represents the angle estimation task loss function, Represents the total number of sample targets, Indicates The probability density of the discrete angles of sample targets in different intervals, Indicates The vector of pose angle labels of sample targets, Indicates The cosine weight of the sample target, , Indicates The pose angle labels of the sample targets.

[0080] For example, when obtaining the distance penalty error of the discrete angle, the distance penalty error of the discrete angle is processed by softmax mapping to obtain the probability density of the discrete angle in different intervals. Since the attitude angle is a variable with inherent periodicity, the present invention uses cosine weighting to process the probability density of the discrete angle interval to obtain a loss function that conforms to the periodicity of the angle. Specifically, The cosine weight of the sample target , when the attitude angle label The cosine value and predicted attitude angle The closer the difference of the cosine values ​​is, The smaller it is, the smaller the loss value corresponding to the sample target is. By controlling the cosine weight to control the loss function of different sample targets, the adaptation to the periodic problem of angle estimation is achieved. The final loss function of the cosine weighted ordered regression is shown in the above formula (8). The loss function of the cosine weighted ordered regression is used as the loss function of the angle estimation task. The form of the cosine weighted ordered regression loss function can be simply regarded as the weighted cross entropy loss, which is used to replace the original mean square error loss function to realize the loss calculation of angle estimation.

[0081] In this embodiment, a cosine weighted ordered regression loss function is used to replace the mean square error loss function originally used for angle estimation, thereby solving the problem of infinite solution space for angle estimation and the periodicity problem of angle estimation.

[0082] In summary, the present invention can use only sample radar HRRP data when the attitude angle is missing during the test process, and use the attitude angle and category as labels to complete the training of the target recognition model, so as to improve the performance of the main task by using auxiliary information; at the same time, the sample selection method of the self-paced strategy is used to design and implement the selection of sample radar HRRP data in the training set during the training process, and the simple samples are trained first through the training idea of ​​"easy first and difficult later", so as to constrain the target recognition model to fit in the direction of the main task and reduce the possibility of falling into the local optimum; the angle estimation task loss function is improved, and the commonly used mean square error loss cannot well describe the error between different attitude angles, because the attitude angle has the characteristics of periodicity. The present invention proposes cosine weighted ordered regression, which improves the estimation effect by discretizing the interval, and at the same time, cosine weighting of each interval can balance the problems caused by periodicity. The AALNet network of the present invention is compared with other single-task networks, and the effect is significantly better than other methods. Table 1 is a performance comparison of the AALNet network of the present invention and other networks. It can be seen from Table 1 that the accuracy of the AALNet network trained based on data set 1 is greater than the accuracy of other networks, the accuracy of the AALNet network trained based on data set 2 is greater than the accuracy of other networks, and when the measured data is input into the AALNet network, the prediction accuracy of the AALNet network is also greater than the prediction accuracy of other networks.

[0083] Table 1

[0084] Among them, AGC-LSTM is an adaptive Gaussian classifier-long short-term memory network (Adaptive GaussianClassifier-Long Short-Term Memory), Deep RF is a deep random forest (DeepRandom Forest), MIConvGRU is a multi-input convolutional gated recurrent unit neural network (Multi-Input Convolutional GatedRecurrent Unit Neural Network), and CNN-Bi-RNN is a convolutional neural network and a bidirectional recurrent neural network (Convolutional Neural Network-Bidirectional Recurrent Neural Network).

[0085] The radar HRRP target recognition device based on attitude angle auxiliary information provided by the present invention is described below. The radar HRRP target recognition device based on attitude angle auxiliary information described below and the radar HRRP target recognition method based on attitude angle auxiliary information described above can be referenced to each other.

[0086] Figure 5 is a schematic diagram of the structure of a radar HRRP target recognition device based on attitude angle auxiliary information provided by an embodiment of the present invention, such as Figure 5 As shown, the radar HRRP target recognition device 500 based on attitude angle auxiliary information includes an acquisition unit 501, an extraction unit 502 and a recognition unit 503; wherein: An acquisition unit 501 is used to acquire radar HRRP data of a target to be identified; An extraction unit 502 is used to input the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; The recognition unit 503 is used to input the feature information into the first specific task network of the target recognition model to obtain the recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0087] The radar HRRP target recognition device based on attitude angle auxiliary information provided by the present invention inputs the acquired radar HRRP data of the target to be recognized into the parameter sharing network of the target recognition model to obtain the feature information output by the parameter sharing network, and then inputs the feature information into the first specific task network of the target recognition model to obtain the recognition result of the target category output by the first specific task network. Wherein, the target recognition model is obtained by training the initial target recognition model based on the sample radar HRRP data of the sample target, the category label of the sample target and the attitude angle label corresponding to the sample target. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle. It can be seen that when training the target recognition model of the present invention, the first specific task network and the second specific task network jointly adopt the output information of the parameter sharing network, and the output information is obtained based on the input sample radar HRRP data, and does not require the input of the attitude angle of the target to be identified. The attitude angle of the target to be identified is only used as a label. Therefore, there is no need to input the attitude angle of the target to be identified in the test phase, thereby avoiding the problem of reduced performance of the target recognition model caused by the lack of attitude angle information in the test phase in the related technology, thereby improving the performance of the target recognition model.

[0088] Based on any of the above embodiments, the target recognition model is trained in the following manner: Acquiring sample radar HRRP data of the sample target; Inputting the sample radar HRRP data into the parameter sharing network of the initial target recognition model to obtain sample feature information output by the parameter sharing network; Inputting the sample feature information into the first specific task network and the second specific task network of the initial target recognition model respectively, to obtain the predicted category output by the first specific task network and the predicted posture angle output by the second specific task network; Determining a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target; Based on the total loss function, the shared parameters of the parameter sharing network, the network parameters of the first specific task network, and the network parameters of the second specific task network are iteratively adjusted to obtain the target recognition model.

[0089] Based on any of the foregoing embodiments, determining a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target includes: Determining a recognition task loss function based on the predicted category and the category label of the sample target; Determining an angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target; The total loss function is determined based on the following formula (1): (1) in, represents the total loss function, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the angle estimation task loss function, Represents the recognition task loss function.

[0090] Based on any of the foregoing embodiments, the acquiring the sample radar HRRP data of the sample target includes: The optimization problem is constructed based on the following formula (2): (2) The constraints are , Represents the total number of sample targets, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the sample selection threshold, Indicates The control parameters of the sample target, It is expressed by the following formula (3): (3) represents the regularization term, , Indicates Multi-task joint loss of sample targets, , represents the weight of the angle estimation task, Indicates The angle estimation task loss function corresponding to the sample target is: Indicates The pose angle labels corresponding to the sample targets are Indicates The predicted posture angle corresponding to the sample target, represents the network parameters of the second task-specific network, Indicates The recognition task loss function corresponding to the sample target is: Indicates Sample targets correspond to classification labels, Indicates The predicted category corresponding to the sample target, representing network parameters of a first task-specific network; Solve the optimization problem to obtain the network parameters of the target recognition model .

[0091] Based on any of the above embodiments, determining the angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target includes: The interval width is determined based on the following formula (4): (4) in, represents the width of the interval, Indicates the maximum angle within the angle space, Indicates the minimum angle within the angle space range, Indicates the total number of pre-set intervals, represents a mapping function; The discrete angle corresponding to the predicted posture angle of the sample target is determined based on the following formula (5): (5) in, represents the predicted attitude angle The corresponding discrete angle; The distance-penalized metric of the discrete angle is determined based on the following formula (6): (6) in, represents the distance penalty metric, represents the serial number of the interval in which the predicted attitude angle is located, represents weight; The discrete angle is determined based on the following formula (7): Distance penalty error relative to different intervals: (7) in, Represents the discrete angle Relative to the distance penalty error of different intervals, Indicates the sequence number of the kth interval; The angle estimation task loss function is determined based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target.

[0092] Based on any of the above embodiments, the determining of the angle estimation task loss function based on the distance penalty error of the discrete angle, the predicted attitude angle, and the attitude angle label corresponding to the sample target includes: Mapping the distance penalty error of the discrete angle to obtain the probability density of the discrete angle in different intervals; The angle estimation task loss function is determined based on the following formula (8): (8) in, represents the angle estimation task loss function, Represents the total number of sample targets, Indicates The probability density of the discrete angles of sample targets in different intervals, Indicates The vector of pose angle labels of sample targets, Indicates The cosine weight of the sample target, , Indicates The pose angle labels of the sample targets.

[0093] Based on any of the above embodiments, the parameter sharing network includes a first convolution module, a second convolution module, a third convolution module and a first fully connected layer connected in sequence, the first convolution module, the second convolution module and the third convolution module all include a convolution layer, a batch normalization layer, a first activation function layer and a maximum pooling layer connected in sequence, and the first specific task network and the second specific task network all include a second fully connected layer, a second activation function layer and a third fully connected layer connected in sequence.

[0094] Figure 6 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the radar HRRP target recognition method based on attitude angle auxiliary information, the method comprising: obtaining radar HRRP data of the target to be recognized; Inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; Inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0095] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0096] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored in a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the radar HRRP target recognition method based on attitude angle auxiliary information provided by the above methods, the method includes: obtaining radar HRRP data of the target to be identified; Inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; Inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0097] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the radar HRRP target recognition method based on attitude angle auxiliary information provided by the above methods, the method comprising: acquiring radar HRRP data of a target to be identified; Inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; Inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

[0098] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0099] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A radar HRRP target recognition method based on attitude angle auxiliary information, characterized in that: include: Obtain radar HRRP data of the target to be identified; Inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; Inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

2. The radar HRRP target recognition method based on attitude angle auxiliary information according to claim 1 is characterized in that: The target recognition model is trained based on the following method: Acquiring sample radar HRRP data of the sample target; Inputting the sample radar HRRP data into the parameter sharing network of the initial target recognition model to obtain sample feature information output by the parameter sharing network; Inputting the sample feature information into the first specific task network and the second specific task network of the initial target recognition model respectively, to obtain the predicted category output by the first specific task network and the predicted posture angle output by the second specific task network; Determining a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target; Based on the total loss function, the shared parameters of the parameter sharing network, the network parameters of the first specific task network, and the network parameters of the second specific task network are iteratively adjusted to obtain the target recognition model.

3. The radar HRRP target recognition method based on attitude angle auxiliary information according to claim 2 is characterized in that: The determining of a total loss function based on the predicted category, the category label of the sample target, the predicted attitude angle, and the attitude angle label corresponding to the sample target includes: Determining a recognition task loss function based on the predicted category and the category label of the sample target; Determining an angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target; The total loss function is determined based on the following formula (1): （1） in, represents the total loss function, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the angle estimation task loss function, Represents the recognition task loss function.

4. The radar HRRP target recognition method based on attitude angle auxiliary information according to claim 2 is characterized in that: The method further comprises: The optimization problem is constructed based on the following formula (2): （2） The constraints are , Represents the total number of sample targets, represents the network parameters of the target recognition model, and represents the trainable parameters, represents the sample selection threshold, Indicates The control parameters of the sample target, It is expressed by the following formula (3): （3） represents the regularization term, , Indicates Multi-task joint loss of sample targets, , represents the weight of the angle estimation task, Indicates The angle estimation task loss function corresponding to the sample target is: Indicates The pose angle labels corresponding to the sample targets are Indicates The predicted posture angle corresponding to the sample target, represents the network parameters of the second task-specific network, Indicates The recognition task loss function corresponding to the sample target is: Indicates Sample targets correspond to classification labels, Indicates The predicted category corresponding to the sample target, representing network parameters of a first task-specific network; Solve the optimization problem to obtain the network parameters of the target recognition model .

5. The radar HRRP target recognition method based on attitude angle auxiliary information according to claim 3 is characterized in that: The determining of the angle estimation task loss function based on the predicted attitude angle and the attitude angle label corresponding to the sample target includes: The interval width is determined based on the following formula (4): （4） in, represents the width of the interval, Indicates the maximum angle within the angle space, Indicates the minimum angle within the angle space range, Indicates the total number of pre-set intervals, represents a mapping function; The discrete angle corresponding to the predicted posture angle of the sample target is determined based on the following formula (5): （5） in, represents the predicted attitude angle The corresponding discrete angle; The distance-penalized metric of the discrete angle is determined based on the following formula (6): （6） in, represents the distance penalty metric, represents the serial number of the interval in which the predicted attitude angle is located, represents weight; The discrete angle is determined based on the following formula (7): Distance penalty error relative to different intervals: （7） in, Represents the discrete angle Relative to the distance penalty error of different intervals, represents the sequence number of the kth interval; The angle estimation task loss function is determined based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target.

6. The radar HRRP target recognition method based on attitude angle auxiliary information according to claim 5 is characterized in that: The determining of the angle estimation task loss function based on the distance penalty error of the discrete angle, the predicted attitude angle and the attitude angle label corresponding to the sample target includes: Mapping the distance penalty error of the discrete angle to obtain the probability density of the discrete angle in different intervals; The angle estimation task loss function is determined based on the following formula (8): （8） in, represents the angle estimation task loss function, Represents the total number of sample targets, Indicates The probability density of the discrete angles of sample targets in different intervals, Indicates The vector of pose angle labels of sample targets, Indicates The cosine weight of the sample target, , Indicates The pose angle labels of the sample targets.

7. The radar HRRP target recognition method based on attitude angle auxiliary information according to any one of claims 1 to 6, characterized in that: The parameter sharing network includes a first convolution module, a second convolution module, a third convolution module and a first fully connected layer connected in sequence, the first convolution module, the second convolution module and the third convolution module all include a convolution layer, a batch normalization layer, a first activation function layer and a maximum pooling layer connected in sequence, and the first specific task network and the second specific task network all include a second fully connected layer, a second activation function layer and a third fully connected layer connected in sequence.

8. A radar HRRP target recognition device based on attitude angle auxiliary information, characterized in that: include: An acquisition unit, used for acquiring radar HRRP data of a target to be identified; An extraction unit, used for inputting the radar HRRP data into a parameter sharing network of a target recognition model to obtain feature information output by the parameter sharing network; A recognition unit, used for inputting the feature information into a first specific task network of the target recognition model to obtain a recognition result of the target category output by the first specific task network; Among them, the target recognition model is obtained by training an initial target recognition model based on sample radar HRRP data of sample targets, category labels of the sample targets and attitude angle labels corresponding to the sample targets. The initial target recognition model includes a parameter sharing network, a first specific task network and a second specific task network. The first specific task network and the second specific task network are both connected to the parameter sharing network. The first specific task network is used to output the predicted category, and the second specific task network is used to output the predicted attitude angle.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the radar HRRP target recognition method based on attitude angle auxiliary information as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the radar HRRP target recognition method based on attitude angle auxiliary information as claimed in any one of claims 1 to 7 is implemented.

Images