CN111352086A - Unknown target identification method based on deep convolutional neural network - Google Patents

Abstract

The invention belongs to the technical field of unknown target recognition, in particular to an unknown target recognition method based on a deep convolutional neural network. The invention firstly preprocesses the one-dimensional range image data (HRRP) obtained by the broadband radar to reduce the amplitude sensitivity of the one-dimensional range image; secondly, the deep convolutional neural network is used to extract features; finally, the difference probability method is used to process the known The recognition probability of the target data is obtained, the discrimination threshold is obtained, and the output vector of the neuron network is discriminated to identify the unknown target. Due to the introduction of the discriminant threshold obtained by the difference probability method, this method effectively describes the statistical distribution area boundary between the known target and the unknown target data set, and solves the problem that the conventional convolutional neural network cannot identify the unknown target.

Description

An Unknown Object Recognition Method Based on Deep Convolutional Neural Networks

technical field

The invention belongs to the technical field of unknown target recognition, in particular to an unknown target recognition method based on a deep convolutional neural network.

Background technique

Since the middle of the last century, the radar target recognition technology has gradually developed and matured, and the target to be recognized by the radar is judged mainly based on the radar target cross-sectional area (RCS) or one-dimensional range profile (HRRP). The high-resolution one-dimensional range is like the vector sum of the target scattering center echo obtained by the broadband radar, which not only provides the geometric shape and structural characteristics of the target, but also contains more relevant information required for target recognition.

In recent years, the theory of deep learning has gradually matured, and convolutional neural networks have been widely used in the field of radar target recognition. Because of its translation insensitivity, nonlinearity and self-learning characteristics, it has achieved good recognition results. However, conventional convolutional neural networks need to be trained on a large amount of data of known targets in advance, which means that conventional convolutional neural networks can only identify targets with known targets (that is, targets that have participated in training). However, in practical applications, it is impossible Acquire the one-dimensional distance image data of all targets in advance, and build a complete convolutional neural network for identification. When the network input is data of unknown targets (that is, targets that do not participate in training), it will be forcibly identified as Knowing the category of the target, leading to misidentification.

SUMMARY OF THE INVENTION

The main content of the present invention is to address the above problems, and propose a method for identifying unknown radar targets based on a deep convolutional neural network. Based on the conventional deep convolutional neural network, the method uses the training one-dimensional distance image data of the known target to obtain the recognition threshold, and effectively describes the statistical distribution area boundary between the known target and the unknown target data set, so as to realize the detection of the unknown target. Recognition solves the problem that conventional neural networks cannot identify unknown targets.

The technical scheme of the present invention is: an unknown target recognition method based on a deep convolutional neural network, comprising the following steps:

S1. Based on the target scattering center model, set the single target one-dimensional range image sample obtained by the broadband radar as x=[x ₁ , x ₂ ,..., _xi ,...,x _N ], where N is the distance The number of units, x _i represents the amplitude of the ith distance unit. In order to reduce the influence of the amplitude sensitivity of the one-dimensional range image on the recognition performance, and to highlight the contrast effect between the strong scattering point and the rest of the scattering points, the one-dimensional range image is subjected to β- Mean normalization:

表示第i个距离单元归一化幅度，β为常数，E_x表示该单幅距离像的均值，β-均值标准化处理后的单幅一维距离像为

in

Represents the normalized amplitude of the i-th distance unit, β is a constant, E _x represents the mean value of the single distance image, and the single one-dimensional distance image after β-mean normalization is

输出为分类器给出的识别标签

每个卷积模块由卷积层、激活函数、批量归一化层、和池化层构成，考虑雷达最小分辨单位以及一维距离像数据特征，卷积核尺寸为1×3，每个卷积层有64个卷积核，池化核为1×2，由于指数线性单元(ELU)具有更好的微分特性，用其代替常用激活函数RELU，提高模型拟合能力，增加对输入变化的鲁棒性。BN层的加入使得训练速度增快，模型快速收敛，激活函数为：S2. Construct a deep convolutional neural network model, and use the improved deep convolutional neural network (DCNN) based on AlexNet to extract high-dimensional features of radar one-dimensional range images under strong supervision learning. The deep convolutional neural network is formed by stacking multiple convolution modules, and adding a dropout layer in the middle makes some neurons randomly inactivated to reduce training parameters and reduce the risk of model overfitting. Based on the principle of neural network, backpropagation (BP) and stochastic gradient descent (SGD) algorithms are used for model convergence and training. As shown in Figure 1, the deep convolutional neural network has a total of 13 layers, followed by convolution module 1, Dropout Layer 1, Convolution Module 2, Convolution Module 3, Dropout Layer 2, Convolution Module 4, Convolution Module 5, Fully Connected Layer 1, Batch Normalization Layer 1, Dropout Layer 3, Fully Connected Layer 2, Batch Normalization Unification layer 2, classifier; the input of the deep convolutional neural network is

The output is the identification label given by the classifier

Each convolution module consists of a convolution layer, an activation function, a batch normalization layer, and a pooling layer. Considering the radar minimum resolution unit and the one-dimensional range image data features, the size of the convolution kernel is 1 × 3, and each volume The product layer has 64 convolution kernels, and the pooling kernel is 1 × 2. Since the exponential linear unit (ELU) has better differential characteristics, it is used instead of the commonly used activation function RELU to improve the model fitting ability and increase the sensitivity to input changes. robustness. The addition of the BN layer makes the training speed faster and the model converges quickly. The activation function is:

S3. Determine the recognition threshold: For conventional convolutional neural network target recognition, it is necessary to obtain the characteristic parameters of all categories during training and learning, and the unknown category target will be forced to be identified as a known category in the classification. In order to identify the unknown type, the present invention introduces the difference probability method to obtain the identification threshold;

In the learning stage of the deep convolutional neural network, the probability output p _ij that the i-th one-dimensional distance image obtained from the classifier belongs to the j-th known category, and the probability vector corresponding to each one-dimensional distance image is p _i =[p _i1 , p _i2 ,...,p _iN ], where N is the number of known categories, select the maximum value p _m and the second maximum value p _sm in the probability vector p _i to obtain the difference probability v _d = p _m -p _sm ;

Enter a single one-dimensional distance image of different categories of targets, and each category of targets can obtain a difference vector:

为第m类目标第i个一维距离像的差值概率，上标T表示转置符；where d ^m is the difference vector of the m-th target,

is the difference probability of the i-th one-dimensional range image of the m-th target, and the superscript T represents the transposition symbol;

Calculate the histogram of the difference probabilities in the difference vector d ^m of all known targets, and select a difference probability from the difference probability histogram as the identification threshold τ according to the predetermined correct discrimination rate of the known targets;

S4. Unknown target recognition:

其中

为第i幅测试一维距离像数据对应的差值概率，M表示未知目标数据个数；Input the obtained single one-dimensional distance image of the unknown target into the trained deep convolutional neural network model to obtain the corresponding difference vector

in

is the difference probability corresponding to the i-th test one-dimensional range image data, and M represents the number of unknown target data;

则将第i幅一维距离像数据识别为已知目标；若差值概率小于门限即

则将第i幅一维距离像数据识别为未知目标，即识别规则为：Compare the difference probability of each one-dimensional range image in d ^t with the recognition threshold τ, if the difference probability is greater than or equal to the threshold, then

Then the i-th one-dimensional range image data is identified as a known target; if the difference probability is less than the threshold, that is

Then the i-th one-dimensional distance image data is identified as an unknown target, that is, the identification rules are:

表示第i幅未知目标一维距离像数据属于已知目标，

表示第i幅未知目标一维距离像数据属于未知目标P。in

Indicates that the i-th unknown target one-dimensional range image data belongs to a known target,

Indicates that the i-th unknown target one-dimensional range image data belongs to the unknown target P.

The beneficial effect of the present invention is that the present invention effectively describes the statistical distribution area boundary between the known target and the unknown target data set due to the introduction of the discrimination threshold obtained by the difference probability method, and solves the problem that the conventional convolutional neural network cannot identify the unknown target. the problem.

Description of drawings

Figure 1 is a schematic diagram of the structure of a deep convolutional neural network model.

Detailed ways

The effectiveness of the present invention is demonstrated below in conjunction with simulation examples.

The simulation one-dimensional range images of five different types of military aircraft, AH64, AN26, F15, B1B, and B52, obtained by the special electromagnetic simulation characteristic scene are used for experiments. The experimental simulation radar parameters include: the radar carrier frequency is 6GHz, and the radar bandwidth is 400MHz. In the simulation scene, the simulation target collects a one-dimensional range image at an elevation angle of 3° and an azimuth angle of 0° to 180° every 0.1°. Each type of aircraft collects 1801 one-dimensional range images, each one-dimensional range image. Each contains 320 distance units, that is, the input data of each type of aircraft is a one-dimensional distance image matrix of 1801 × 320.

In the process of training and updating the parameter W, randomly initialize the weight W=[w ₁ ,w ₂ ,w ₃ ] and the bias B=[b ₁ ,b ₂ ,...,b _N ], and select the cross entropy loss function as loss function and parameters optimized by the Adam optimizer with an adaptive learning rate initialized to 0.0001.

The recognition results of the above five types of simulated radar target data using the conventional convolutional neural network method and the method in this paper are shown in Table 1:

Table 1 The recognition results of the two methods for unknown targets

It can be seen from the experimental results that when three types of aircraft are randomly selected as known targets and the other two types of aircraft are used as unknown targets, the conventional CNN network cannot identify the unknown target, and the method in this paper introduces unknown target discrimination. threshold, the unknown target can be well identified, and the average correct recognition rate for the unknown target is above 80%, thus verifying that the method of the present invention is effective.

Claims (1)

1. an unknown target recognition method based on deep convolutional neural network, is characterized in that, comprises the following steps: S1. Based on the target scattering center model, set the single target one-dimensional range image sample obtained by the broadband radar as x=[x ₁ ,x ₂ ,...,x _i ,...,x _N ], where N is the distance unit number number, x _i represents the magnitude of the ith distance unit, and β-mean normalization is performed on the one-dimensional distance image:

in

Represents the normalized amplitude of the i-th distance unit, β is a constant, E _x represents the mean value of the single distance image, and the single one-dimensional distance image after β-mean normalization is

S2. Build a deep convolutional neural network model. The deep convolutional neural network has a total of 13 layers, which are convolution module 1, Dropout layer 1, convolution module 2, convolution module 3, Dropout layer 2, convolution module 4, Convolution module 5, fully connected layer 1, batch normalization layer 1, Dropout layer 3, fully connected layer 2, batch normalization layer 2, classifier; the input of the deep convolutional neural network is

The output is the identification label given by the classifier

Each convolution module consists of a convolution layer, an activation function, a batch normalization layer, and a pooling layer. The size of the convolution kernel is 1×3, and each convolution layer has 64 convolution kernels and pooling kernels. is 1×2, and the activation function is:

S3. Determine the recognition threshold: in the learning stage of the deep convolutional neural network, the probability output p _ij that the i-th one-dimensional distance image obtained from the classifier belongs to the j-th known category, and each one-dimensional distance image corresponds to The probability vector of is p _i =[p _i1 ,p _i2 ,...,p _iN ], where N is the number of known categories, select the maximum value p _m and the second maximum value p _sm in the probability vector p _i to obtain Its difference probability v _d =p _{m -psm} ; Enter a single one-dimensional distance image of different categories of targets, and each category of targets can obtain a difference vector:

where d ^m is the difference vector of the m-th target,

is the difference probability of the i-th one-dimensional range image of the m-th target, and the superscript T represents the transposition symbol; Calculate the histogram of the difference probabilities in the difference vector d ^m of all known targets, and select a difference probability from the difference probability histogram as the identification threshold τ according to the predetermined correct discrimination rate of the known targets; S4. Unknown target recognition: Input the obtained single one-dimensional distance image of the unknown target into the trained deep convolutional neural network model to obtain the corresponding difference vector

in

is the difference probability corresponding to the i-th test one-dimensional range image data, and M represents the number of unknown target data; Compare the difference probability of each one-dimensional range image in d ^t with the recognition threshold τ, if the difference probability is greater than or equal to the threshold, then

Then the i-th one-dimensional range image data is identified as a known target; if the difference probability is less than the threshold, that is

Then the i-th one-dimensional distance image data is identified as an unknown target, that is, the identification rules are:

in

Indicates that the i-th unknown target one-dimensional range image data belongs to a known target,

Indicates that the i-th unknown target one-dimensional distance image data belongs to the unknown target.

Images