Object detection method and device based on deep convolutional feature matrix manifold network

Abstract

The present application relates to a target detection method and device based on a deep convolution feature matrix manifold network. The method comprises: using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to obtain echo data of a simulated target with different signal-to-noise ratios; obtaining a time-frequency feature map of the echo data using a time-frequency analysis method; extracting high-dimensional nonlinear features of the time-frequency feature map using a pre-trained convolutional neural network to obtain a deep convolution feature matrix; and training a pre-constructed deep manifold network using the deep convolution feature matrix as a sample, so that the deep manifold network can detect targets in the echo data to be detected. The use of this method can improve radar target detection performance.

Description

Technical Field

[0001] The present application relates to the field of radar target detection technology, and in particular to a target detection method and device based on a deep convolutional feature matrix manifold network. Background Art

[0002] Radar systems are a core component of modern surveillance and defense systems and are widely used in fields such as aviation, aerospace, and weather forecasting. The primary challenge in radar target detection lies in accurately detecting and identifying targets amidst complex environmental noise and interference. Clutter, or non-target reflected signals received by the radar system, is a major source of interference that affects radar detection performance. Different environments, such as ocean surfaces, mountainous areas, and urban areas, generate different types of clutter, each impacting radar target detection to varying degrees.

[0003] Traditional detection technology based on statistical characteristics has high requirements on the target signal-to-noise ratio, slow detection speed, many false alarms and model mismatch problems, and its detection performance cannot meet the needs of modern complex and diverse perception environments. Summary of the Invention

[0004] Based on this, it is necessary to provide a target detection method and device based on a deep convolutional feature matrix manifold network to address the above technical problems.

[0005] A target detection method based on a deep convolutional feature matrix manifold network, the method comprising:

[0006] Using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to obtain echo data of simulated targets having different signal-to-noise ratios; the Doppler steering vector is represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient;

[0007] A time-frequency feature map of the echo data is obtained by using a time-frequency analysis method, and a pre-trained convolutional neural network is used to extract high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix;

[0008] The deep convolution feature matrix is ​​used as a sample to train a pre-built deep manifold network, so that the deep manifold network can detect the target in the echo data to be detected.

[0009] In one embodiment, the method further includes: using Doppler steering vector driven simulated target signals in pre-acquired sea clutter data to be:

[0010]

[0011] α=β×mean(S)cosθ

[0012] Where S is the sea clutter data, s is the Doppler steering vector, α is the target signal amplitude parameter, θ represents the initial phase, β represents the scattering coefficient, N represents the number of target signal pulses, and f d represents the normalized Doppler frequency.

[0013] In one embodiment, the method further includes: obtaining a time-frequency characteristic diagram of the echo data by using an STFT time-frequency analysis method.

[0014] In one embodiment, the method further includes: inputting the time-frequency feature map into a pre-trained convolutional neural network, and using the convolution layer output of the convolutional neural network as a high-dimensional nonlinear convolution feature; symmetric positive-definitely performing the averaging on the matrices of all channels of the high-dimensional nonlinear convolution feature to obtain a deep convolution feature matrix.

[0015] In one embodiment, the method further includes: symmetric positive definite transformation of matrices of all channels of the high-dimensional nonlinear convolution feature into:

[0016]

[0017] Wherein, X0 represents the matrix of each channel in the high-dimensional nonlinear convolution feature, X2 represents the high-dimensional nonlinear convolution feature after positive definiteness, ε represents the correction threshold, I is the unit matrix, and U1 represents the orthogonal matrix;

[0018] The average value of the positively definite high-dimensional nonlinear convolution features is:

[0019]

[0020] Among them, S1(i,j,k) represents the high-dimensional nonlinear convolution feature with k channels after positive definiteness, and S2(i,j) represents the deep convolution feature matrix.

[0021] In one embodiment, the number of convolutional layers of the convolutional neural network is 1 or 2.

[0022] In one embodiment, when the number of convolutional layers of the convolutional neural network is 1, the number of network layers of the deep manifold network when the signal-to-noise ratio is greater than the first signal-to-noise ratio is 4, the number of network layers when the signal-to-noise ratio is less than the first signal-to-noise ratio and greater than the second signal-to-noise ratio is 8, and the number of network layers when the signal-to-noise ratio is less than the second signal-to-noise ratio is 6; when the number of convolutional layers of the convolutional neural network is 2, the number of network layers of the deep manifold network is 8.

[0023] A target detection device based on a deep convolutional feature matrix manifold network, the device comprising:

[0024] a data acquisition module configured to acquire echo data of simulated targets having different signal-to-noise ratios using a simulated target signal driven by a Doppler steering vector from pre-acquired sea clutter data; the Doppler steering vector being represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient;

[0025] A deep convolution feature matrix construction module is used to obtain a time-frequency feature map of the echo data using a time-frequency analysis method, and to extract high-dimensional nonlinear features of the time-frequency feature map using a pre-trained convolutional neural network to obtain a deep convolution feature matrix;

[0026] The target detection module is used to train a pre-built deep manifold network using the deep convolution feature matrix as a sample, so that the deep manifold network can detect targets in the echo data to be detected.

[0027] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0028] Using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to obtain echo data of simulated targets having different signal-to-noise ratios; the Doppler steering vector is represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient;

[0029] A time-frequency feature map of the echo data is obtained by using a time-frequency analysis method, and a pre-trained convolutional neural network is used to extract high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix;

[0030] The deep convolution feature matrix is ​​used as a sample to train a pre-built deep manifold network, so that the deep manifold network can detect the target in the echo data to be detected.

[0031] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0032] Using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to obtain echo data of simulated targets having different signal-to-noise ratios; the Doppler steering vector is represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient;

[0033] A time-frequency feature map of the echo data is obtained by using a time-frequency analysis method, and a pre-trained convolutional neural network is used to extract high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix;

[0034] The deep convolution feature matrix is ​​used as a sample to train a pre-built deep manifold network, so that the deep manifold network can detect the target in the echo data to be detected.

[0035] The above-mentioned target detection method and device based on the deep convolution feature matrix manifold network use the simulated target signal driven by the Doppler steering vector in the pre-acquired sea clutter data to obtain the echo data of the simulated target with different signal-to-noise ratios, use the time-frequency analysis method to obtain the time-frequency feature map of the echo data, use the pre-trained convolutional neural network to extract the high-dimensional nonlinear features of the time-frequency feature map, obtain the deep convolution feature matrix, and use the deep convolution feature matrix as a sample to train the pre-constructed deep manifold network so that the deep manifold network can detect the target in the echo data to be detected. The present invention proposes to use the convolutional neural network to first extract the high-dimensional nonlinear features from the feature map, and then use the deep matrix manifold network to reduce the dimensionality and aggregate it to expand more features, thereby improving the radar target detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 1 is a flow chart of an object detection method based on a deep convolutional feature matrix manifold network in one embodiment;

[0037] Figure 2 1 is a diagram of sea clutter data in one embodiment; wherein (a) is a diagram of nearby sea clutter data, and (b) is a diagram of selected sea clutter data;

[0038] Figure 3 1 is a signal diagram after adding a simulation target in one embodiment; (a) is a signal diagram after adding a 0dB simulation target, and (b) is a signal diagram after adding a -5dB simulation target;

[0039] Figure 4 This is a diagram of the SPDNet structure in one embodiment;

[0040] Figure 5 : is an STFT characteristic diagram of the signal-to-noise ratio in one embodiment; wherein (a) is no target, (b) is 0 dB, (c) is -5 dB, (d) is -10 dB, (e) is -15 dB, and (f) is -20 dB;

[0041] Figure 6 : is the STFT characteristic matrix of the signal-to-noise ratio in one embodiment; wherein (a) is no target, (b) is 0 dB, (c) is -5 dB, (d) is -10 dB, (e) is -15 dB, and (f) is -20 dB;

[0042] Figure 7 This is a flow chart of convolutional feature reconstruction in one embodiment;

[0043] Figure 8 Schematic diagram of selecting the signal-to-noise ratio of a training fusion network in one embodiment; wherein (a) represents fusion network 1, and (b) represents fusion network 2;

[0044] Figure 9 Detection probability graph of different networks in one embodiment;

[0045] Figure 10 Detection probability diagrams of the fusion network under different learning rates in one embodiment; wherein (a) represents fusion network 1, and (b) represents fusion network 2;

[0046] Figure 11 Detection probability diagrams of fusion networks under different weight attenuation in one embodiment; (a) represents fusion network 1, and (b) represents fusion network 2;

[0047] Figure 12 Detection probability diagrams of fused networks with different numbers of network layers in one embodiment; (a) represents fused network 1, and (b) represents fused network 2;

[0048] Figure 13 1 is a block diagram of a target detection device based on a deep convolutional feature matrix manifold network in one embodiment;

[0049] Figure 14 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0051] In one embodiment, Figure 1 As shown, a target detection method based on a deep convolutional feature matrix manifold network is provided, comprising the following steps:

[0052] Step 102 : Using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data, to acquire echo data of a simulated target having different signal-to-noise ratios.

[0053] The Doppler steering vector is represented by the target signal amplitude parameter, and the size of the target signal amplitude parameter is related to the average size of the sea clutter data, signal-to-noise ratio, initial phase and scattering coefficient.

[0054] Specifically, the radar of the present invention is an X-band solid-state power amplifier surveillance / navigation radar with a transmission time of 40ns-100us and a 360° omnidirectional scanning in the horizontal plane. The radar technical parameters are shown in Table 1.

[0055] Table 1 Radar parameters

[0056]

[0057]

[0058] The present invention uses the "20191012112446_01_staring.mat" in the public data to construct a network dataset, which contains 10,000 pulse trains and a total of 5,250 distance units of sea clutter data. The radar echo data are all pulse compressed data. In the present invention, the 186th to 190th distance units are selected as experimental clutter sample data, such as Figure 2 (a) Intercept the distance unit near the data selected by the present invention, where the data in the red box is the experimental data selected by the present invention. Figure 2 (b) shows the sea clutter data after selecting the range unit.

[0059] In this step, the signal-to-noise ratio is a variable in the range of [0dB, -20dB], the initial phase θ of the target is 0, and the scattering coefficient β is 0.8. The N of the simulation data set is 10000, and f d Set to 0.15.

[0060] Figure 3 Figures (a) and (b) show the signal plots after adding 0dB and -5dB simulated targets. It's clear that adding a high SNR target significantly alters the data plot, while adding a relatively low SNR signal barely changes the data plot. Visually, the clutter and target signal plots are virtually indistinguishable. Because the added steering vector is a phase factor, the specific location of the signal cannot be directly observed on the signal amplitude plot.

[0061] Step 104: Obtain a time-frequency feature map of the echo data using a time-frequency analysis method, and use a pre-trained convolutional neural network to extract high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix.

[0062] In this step, for the matrix A n×n Diagonalization and eigenvalue decomposition can be performed:

[0063] U -1 AU=Σ

[0064] A=U -1 ΣU

[0065] The eigenvalues ​​and eigenvectors of a square matrix are the essential characteristics of a matrix. m×n It is impossible to perform eigenvalue decomposition, so the singular value decomposition method is used to decompose the non-square matrix B m×n Decompose. Since the singular value decomposition originates from the nature of the matrix, it is applicable to all real matrices:

[0066] BV=UΣ

[0067] B=UΣV -1 =UΣV T

[0068] Among them, U is an m×m orthogonal matrix, Σ is an m×n orthogonal matrix, and V is an n×n orthogonal matrix. The symmetric positive definite method of any matrix X0 is to transpose the matrix X0 and its conjugate When multiplied, the generated matrix X1 must have conjugate symmetry, and then by correcting the negative eigenvalues, the time-frequency characteristic matrix is ​​obtained.

[0069] Step 106 : Using the deep convolution feature matrix as a sample, training a pre-built deep manifold network so that the deep manifold network can detect the target in the echo data to be detected.

[0070] In the above-mentioned target detection method based on the deep convolution feature matrix manifold network, a Doppler steering vector-driven simulated target signal is used in the pre-acquired sea clutter data to obtain echo data of the simulated target with different signal-to-noise ratios, and the time-frequency feature map of the echo data is obtained by time-frequency analysis. The high-dimensional nonlinear features of the time-frequency feature map are extracted using a pre-trained convolutional neural network to obtain a deep convolution feature matrix. The deep convolution feature matrix is ​​used as a sample to train a pre-constructed deep manifold network so that the deep manifold network can detect the target in the echo data to be detected. The present invention proposes to use a convolutional neural network to first extract high-dimensional nonlinear features from the feature map, and then use a deep matrix manifold network to reduce and aggregate the dimensionality to expand more features, thereby improving the radar target detection performance.

[0071] It's worth noting that a deep manifold network is a deep learning algorithm that processes matrix manifolds. It uses matrix data as input to a neural network for nonlinear training on the matrix manifold. Among these networks, the deep manifold network (SPDNet) is a commonly used architecture.

[0072] Figure 4This is the structural diagram of SPDNet. The generated SPD data is fed into the input layer. After being processed through multiple bilinear mapping layers and feature correction layers, the dimension of the SPD matrix gradually decreases and the features gradually aggregate. Ultimately, these matrices are mapped to Euclidean space using the log-Euclidean metric and classified in space. The original network in this invention has six dimensionality reduction and aggregation iteration layers, namely three bilinear mapping layers and three feature correction layers.

[0073] Similar to classic convolutional neural networks, SPDNet consists of two key functional layers: the Bilinear Map Layer (BiMap Layer) and the Eigenvalue Rectification Layer (ReEig Layer). After processing by these layers, the original SPD matrix is ​​transformed into a new matrix with smaller dimensions and more distinct features. This transformation enhances the discrimination between targets and clutter on the matrix manifold.

[0074] The purpose of the bilinear mapping layer is to make the input SPD matrix smaller and more discriminative. Through the bilinear mapping layer, the input on the original SPD manifold is converted to a new input, forming a new SPD manifold. In other words, the data space on each bilinear mapping layer corresponds to an SPD manifold. The calculation formula of the bilinear mapping layer is defined as follows:

[0075]

[0076] Among them, X k-1 is the input matrix of the k-1th layer, X k is the output matrix of the kth layer, W k is a weight matrix with full row rank.

[0077] The purpose of the feature correction layer is to correct the input SPD matrix, correcting the negative eigenvalue to a positive number, ensuring the symmetric positive definiteness of the matrix, and thus retaining more significant features. The calculation formula of the eigenvalue correction layer is defined as follows:

[0078]

[0079] Among them, U k-1 and Σ k-1 By eigenvalue decomposition Obtain, ε is the correction threshold, and I is the identity matrix.

[0080] After passing through multiple bilinear mapping layers and eigenvalue correction layers, the SPD matrix enters the Log-Euclidean calculation layer (Eigenvalue Logarithm, LogEig Layer). The logarithmic Euclidean metric gives the SPD matrix manifold a Lie group structure that follows the characteristics of the Riemannian manifold and expands the SPD matrix manifold into a flat Euclidean space to form a vector

[48] . The calculation formula of the Log-Euclidean calculation layer is as follows:

[0081]

[0082] Among them, log(Σ k-1 ) is a diagonal matrix composed of logarithmic eigenvalues. After logarithmic Euclidean metric processing, the linear classification method is used to classify the input SPD sample matrix.

[0083] In classification problems, the final output layer is generally a Softmax layer. The vectors that pass through the LogEig layer are passed through the Softmax function, which converts the output values ​​of the multi-classification into a probability distribution in the range [0, 1] with a sum of 1. The calculation formula of Softmax is as follows:

[0084]

[0085] Among them, x i is the output of the i-th node, C is the number of output nodes, that is, the number of classifications, here is the probability of having a target or not having a target, and the final detection probability is obtained from the classification result.

[0086] In addition to the forward propagation process, SPDNet's backpropagation process uses the Stochastic Gradient Descent (SGD) algorithm. This method is a simple but very effective method that can be used for both classification and regression calculations. SGD randomly extracts a group of samples from the training set, updates them once according to the gradient, then extracts another group and updates them again. The randomness here means that the samples are randomly shuffled during each iteration to effectively reduce the parameter update offset problem caused by different samples. The calculation formula is as follows:

[0087]

[0088] Among them, α is the network learning rate, θ j is the jth coefficient of the prediction function, x(i) is the input data, y(i) is the actual value of each sample in the data set, h θ is the prediction function, which calculates the predicted y(i) value based on each input x(i) according to θ.

[0089] In one embodiment, the simulated target signal driven by Doppler steering vector in pre-acquired sea clutter data is:

[0090]

[0091] α=β×mean(S)cosθ

[0092] Where S is the sea clutter data, s is the Doppler steering vector, α is the target signal amplitude parameter, θ represents the initial phase, β represents the scattering coefficient, N represents the number of target signal pulses, and f d represents the normalized Doppler frequency.

[0093] In another embodiment, the STFT time-frequency analysis method is used to obtain the time-frequency characteristic diagram of the echo data.

[0094] STFT time-frequency analysis method:

[0095] Specifically, after superimposing the radar echo data of each range unit with simulated targets of different signal-to-noise ratios, a Hamming window with a length of 25, an overlap length of 24, and 100 sampling points are used to perform short-time Fourier transform. The size of the generated time-frequency graph is 100×76. Randomly taking the 77th time-frequency graph of the 187th range unit as an example, the following is obtained: Figure 5 Feature maps of the signal-to-noise ratio of the parts without and with targets.

[0096] Observe the time-frequency characteristic diagram with and without targets. From the perspective of amplitude, the clutter is concentrated in the low-frequency part, and the clutter spectrum line distribution amplitude is large. The target is distributed at the frequency f d = 0.15, the amplitude decreases as the signal-to-noise ratio decreases. The graphical features show that the clutter exhibits an irregular distribution, with distributions across all frequencies, while the time-frequency distribution of the target is a straight line. The STFT experimental results show a corresponding relationship between the target and clutter frequency distributions.

[0097] For feature maps with different signal-to-noise ratios, the target's time-frequency characteristic amplitude decreases significantly as the signal-to-noise ratio decreases. When the signal-to-noise ratio drops to -10dB, only a small amount of signal can be seen. As the signal-to-noise ratio continues to decrease, it is no longer possible to distinguish the target from the clutter in the image with the naked eye.

[0098] The obtained time-frequency graph is symmetrically positive-definite and converted into a time-frequency feature matrix with a size of 100×100, such as Figure 6 shown.

[0099] The STFT feature matrix shows that the center point on the matrix diagonal represents low-frequency clutter, while the point in the upper left represents the target. Corresponding to the feature map, when the signal-to-noise ratio decreases to -10dB, the target can still be distinguished from the image color depth. However, as the signal-to-noise ratio decreases further, it becomes difficult to distinguish the target from the clutter in the feature matrix.

[0100] In one embodiment, the time-frequency feature map is input into a pre-trained convolutional neural network, and the convolution layer output of the convolutional neural network is used as a high-dimensional nonlinear convolution feature; the matrices of all channels of the high-dimensional nonlinear convolution feature are symmetric-normalized and then averaged to obtain a deep convolution feature matrix.

[0101] Specifically, the convolutional pooling layer in a convolutional neural network extracts high-dimensional, nonlinear features from the target. Similar structures in a deep matrix manifold network—the bilinear mapping layer and feature correction layer—simply reduce the dimensionality of the matrix to remove redundant information. However, for small, weak targets, simply reducing and aggregating the features will not lead to superior detection performance. Therefore, we propose using a convolutional neural network to first extract high-dimensional, nonlinear features from the feature map, and then using a deep matrix manifold network to reduce and aggregate the features to expand the features.

[0102] The convolution module of any convolutional neural network can extract high-dimensional nonlinear features. In order to carry out the experiment of the proposed method, GoogleNet with fewer network parameters is taken as an example. GoogleNet has two convolutional pooling modules before entering the Inception layer. The corresponding input and output sizes are shown in Table 2. To facilitate the research, the trained GoogleNet is used to perform preliminary feature extraction on the generated STFT time-frequency map through transfer learning. The output of these two convolutional pooling layers is used as the extracted high-dimensional nonlinear convolutional features. The process is as follows: Figure 7 As shown in Figure 3, the convolutional feature extraction module is also applicable to other convolutional neural networks.

[0103] Table 2 GoogleNet convolutional layer input and output sizes

[0104]

[0105] Among them, fusion network 1 refers to the process of directly putting the features into SPDNet after the features are extracted by the Pool_1 layer, while fusion network 2 refers to the process of extracting features by the Pool_1 layer and Pool_2 layer and then putting them into SPDNet for processing. The difference between the two is a convolution pooling module.

[0106] The sample data selected in the experiment are all the extracted STFT time-frequency feature maps as the target of convolution feature extraction. After the convolution operation of fusion network 1 and fusion network 2, all the time-frequency feature maps are extracted as convolution feature matrices.

[0107] For fusion network 1, after training on multiple signal-to-noise ratio data, the fusion network 1 trained on the target and clutter dataset with a signal-to-noise ratio of -7dB has the best detection performance, such as Figure 8 As shown in (a). For fusion network 2, the detection performance of the target and clutter dataset with a signal-to-noise ratio of -8dB is the best, as shown in Figure 8 (b) As the SNR continues to increase, the network still fails to converge and training ceases. This experimental phenomenon proves that the proposed method can break through the training limit of -5dB SNR samples, thereby improving the network's detection limit for weak targets.

[0108] The convolution feature matrix corresponding to the optimal network signal-to-noise ratio is classified into 40% training samples and 20% validation samples. The remaining 40% test samples of all signal-to-noise ratio data are used to test the detection probability distribution of fusion network 1 and fusion network 2.

[0109] Because the output of the convolutional layer is a feature matrix with a certain channel depth, the three-dimensional feature depth needs to be compressed into a two-dimensional feature map. After the input image enters the convolutional neural network, although the convolution kernel of each layer is the same, the number of convolution calculations varies, resulting in different feature maps. The final output is a superposition of the feature maps of all channels. The matrix of each channel output of the convolutional layer is symmetric and positively defined, and then the sum and average of the channel matrices are taken, so that the three-dimensional feature matrix of size 57×57×64 or 57×57×192 is compressed to a two-dimensional feature matrix of size 57×57.

[0110] In one embodiment, the matrix of all channels of the high-dimensional nonlinear convolution feature is symmetric and positively defined as follows:

[0111]

[0112] Among them, X0 represents the matrix of each channel in the high-dimensional nonlinear convolution feature, X2 represents the high-dimensional nonlinear convolution feature after positive definiteness, ε represents the correction threshold, I is the unit matrix, and U1 represents the orthogonal matrix; the average value of the high-dimensional nonlinear convolution feature after positive definiteness is:

[0113]

[0114] Among them, S1(i,j,k) represents the high-dimensional nonlinear convolution feature with k channels after positive definiteness, and S2(i,j) represents the deep convolution feature matrix.

[0115] Among them, S1 is the three-dimensional feature matrix with channel depth, S2 is the processed two-dimensional feature matrix, and k is the channel depth.

[0116] From the feature matrix diagram of fusion network 1, we can see that the center point on the matrix diagonal is low-frequency clutter, and the other points on the diagonal are targets. However, because the generated feature map is too abstract after convolution extraction, when the signal-to-noise ratio is less than -10dB, it is difficult to clearly distinguish between clutter and targets with the naked eye.

[0117] The feature matrix plot for Fusion Network 2, which has a higher signal-to-noise ratio, shows that clutter is located at the center of the matrix, while target features are primarily distributed on either side of the central diagonal. The feature matrix for Fusion Network 2, after adding an additional convolutional layer, exhibits a greater average depth than Fusion Network 1, exhibiting more abstract features. Even at higher signal-to-noise ratios, it is difficult to distinguish between target and clutter.

[0118] In one embodiment, the number of convolutional layers of the convolutional neural network is 1 or 2.

[0119] In one embodiment, the number of network layers of the deep manifold network is 4 when the signal-to-noise ratio is greater than the first signal-to-noise ratio, the number of network layers is 8 when the signal-to-noise ratio is less than the first signal-to-noise ratio and greater than the second signal-to-noise ratio, and the number of network layers is 6 when the signal-to-noise ratio is less than the second signal-to-noise ratio; when the number of convolutional layers of the convolutional neural network is 2, the number of network layers of the deep manifold network is 8.

[0120] In summary, based on the above method, the results are analyzed:

[0121] The initial network training hyperparameters of fusion network 1 and fusion network 2 are shown in Table 3. A larger learning rate is used in the training of the two fusion networks to ensure that the fusion network can converge better with the trained SPDNet when other parameters are the same.

[0122] Table 3. Fusion network model training hyperparameters

[0123]

[0124] from Figure 9 As can be seen from the figure, both fusion network 1 and fusion network 2 achieve significantly better target detection performance than SPDNet. When the detection probability is 70%, the corresponding signal-to-noise ratio of fusion network 1 is -9.46dB, that of fusion network 2 is -9.3dB, and that of SPDNet is only -8.12dB. Therefore, under the same conditions, fusion network 2 improves detection performance by 1.34dB over SPDNet. For both fusion networks, due to the additional convolutional pooling process, the time-frequency features are further extracted and aggregated, resulting in a 0.16dB improvement in detection performance for fusion network 2 over fusion network 1.

[0125] The experimental results show that both fusion network 1 and fusion network 2, combined with the ability of convolutional neural networks to extract high-dimensional nonlinear features, increase the feature information in the features and break through the lower limit of the signal-to-noise ratio of target detection through dimensionality reduction aggregation alone. It proves that the detection performance of the proposed fusion network method is better than that of the deep manifold network SPDNet.

[0126] Learning rate:

[0127] Under the condition of ensuring the convergence of the fusion network training loss function, the influence of the learning rate on the fusion network detection performance is studied. Figure 10 As shown in the figure, for fusion network 1, a learning rate of 0.01 achieves the best detection performance under all signal-to-noise ratios. A learning rate that is too large will cause the network to overfit, while a learning rate that is too small may cause the network to underfit. Both will reduce the detection performance of fusion network 1. For fusion network 2, when the signal-to-noise ratio is greater than -9dB, the detection performance is better when the learning rate is 0.01. When the signal-to-noise ratio is less than -9dB, the detection performance is better when the learning rate is 0.005. A learning rate of 0.02 will have a certain overfitting problem, so the detection performance is not as good as a smaller learning rate.

[0128] Weight decay:

[0129] like Figure 11 As shown in the figure, for fusion network 1, the detection performance is the best under all signal-to-noise ratios when the weight decay is 0.0005. However, for fusion network 2, the network training does not produce overfitting, and the network training effect is better.

[0130] Number of network layers:

[0131] This chapter selects comparative experiments with 4, 6, 8, and 10 network layers to study the ideal number of network layers for fusion network 1 and fusion network 2, ensuring that the final dimensionality reduction aggregation matrix size of all network layers is the same. The dimensionality reduction matrix size corresponding to the network layer number is shown in Table 3.

[0132] Table 4.3 Fusion network layer number reduction matrix

[0133]

[0134] like Figure 12 As shown, for fusion network 1, the performance difference between different network layers is small, but when the signal-to-noise ratio is greater than -7.67dB, the detection performance of network layer 4 is the best. When the signal-to-noise ratio is less than -8dB and greater than -7.61dB, the detection performance of network layer 8 is the best. When the signal-to-noise ratio is less than -8dB, the detection performance of network layer 6 is the best. For fusion network 2, the detection performance of network layer 8 is the best under all signal-to-noise ratios.

[0135] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in the present invention, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0136] In one embodiment, Figure 13 As shown, a target detection device based on a deep convolution feature matrix manifold network is provided, including: a data acquisition module 1302, a deep convolution feature matrix construction module 1304 and a target detection module 1306, wherein:

[0137] The data acquisition module 1302 is configured to use a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to acquire echo data of a simulated target having different signal-to-noise ratios; the Doppler steering vector is represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient;

[0138] A deep convolution feature matrix construction module 1304 is configured to obtain a time-frequency feature map of the echo data using a time-frequency analysis method, and extract high-dimensional nonlinear features of the time-frequency feature map using a pre-trained convolutional neural network to obtain a deep convolution feature matrix.

[0139] The target detection module 1306 is used to train a pre-built deep manifold network using the deep convolution feature matrix as a sample, so that the deep manifold network can detect targets in the echo data to be detected.

[0140] In one embodiment, the data acquisition module 1302 is further configured to use the simulated target signal driven by the Doppler steering vector in the pre-acquired sea clutter data as follows:

[0141]

[0142] α=β×mean(S)cosθ

[0143] Where S is the sea clutter data, s is the Doppler steering vector, α is the target signal amplitude parameter, θ represents the initial phase, β represents the scattering coefficient, N represents the number of target signal pulses, and f d represents the normalized Doppler frequency.

[0144] In one embodiment, the deep convolution feature matrix construction module 1304 is further configured to obtain a time-frequency feature map of the echo data using an STFT time-frequency analysis method.

[0145] In one embodiment, the deep convolution feature matrix construction module 1304 is further used to input the time-frequency feature map into a pre-trained convolutional neural network, and use the convolution layer output of the convolutional neural network as a high-dimensional nonlinear convolution feature; and calculate the average value of the matrices of all channels of the high-dimensional nonlinear convolution feature after symmetric positive definiteness to obtain a deep convolution feature matrix.

[0146] In one embodiment, the deep convolution feature matrix construction module 1304 is further configured to symmetric positively define the matrices of all channels of the high-dimensional nonlinear convolution feature as follows:

[0147]

[0148] Among them, X0 represents the matrix of each channel in the high-dimensional nonlinear convolution feature, X2 represents the high-dimensional nonlinear convolution feature after positive definiteness, ε represents the correction threshold, I is the unit matrix, and U1 represents the orthogonal matrix; the average value of the high-dimensional nonlinear convolution feature after positive definiteness is:

[0149]

[0150] Among them, S1(i,j,k) represents the high-dimensional nonlinear convolution feature with k channels after positive definiteness, and S2(i,j) represents the deep convolution feature matrix.

[0151] In one embodiment, the number of convolutional layers of the convolutional neural network is 1 or 2.

[0152] In one embodiment, when the number of convolutional layers of the convolutional neural network is 1, the number of network layers of the deep manifold network when the signal-to-noise ratio is greater than the first signal-to-noise ratio is 4, the number of network layers when the signal-to-noise ratio is less than the first signal-to-noise ratio and greater than the second signal-to-noise ratio is 8, and the number of network layers when the signal-to-noise ratio is less than the second signal-to-noise ratio is 6; when the number of convolutional layers of the convolutional neural network is 2, the number of network layers of the deep manifold network is 8.

[0153] For the specific definition of the target detection device based on the deep convolution feature matrix manifold network, please refer to the definition of the target detection method based on the deep convolution feature matrix manifold network mentioned above, which will not be repeated here. The various modules in the above-mentioned target detection device based on the deep convolution feature matrix manifold network can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0154] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 14 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a target detection method based on a deep convolution feature matrix manifold network is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0155] Those skilled in the art will understand that Figure 14 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0156] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the method in the above embodiment when executing the computer program.

[0157] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiment are implemented.

[0158] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0159] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0160] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A target detection method based on deep convolutional feature matrix manifold network, characterized in that: The method comprises: Using a simulated target signal driven by a Doppler steering vector in pre-acquired sea clutter data to obtain echo data of simulated targets having different signal-to-noise ratios; the Doppler steering vector is represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient; A time-frequency feature map of the echo data is obtained by using a time-frequency analysis method, and a pre-trained convolutional neural network is used to extract high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix; Using the deep convolution feature matrix as a sample, training a pre-built deep manifold network so that the deep manifold network detects targets in the echo data to be detected; A pre-trained convolutional neural network is used to extract the high-dimensional nonlinear features of the time-frequency feature map to obtain a deep convolution feature matrix, including: Inputting the time-frequency feature map into a pre-trained convolutional neural network, and using the convolutional layer output of the convolutional neural network as a high-dimensional nonlinear convolution feature; The matrices of all channels of the high-dimensional nonlinear convolution feature are symmetrically positive-defined and then averaged to obtain a deep convolution feature matrix; The matrices of all channels of the high-dimensional nonlinear convolution feature are symmetrically positive-defined and averaged to obtain a deep convolution feature matrix, including: The matrix of all channels of the high-dimensional nonlinear convolution feature is symmetric and positively defined as follows: in, Represents the matrix of each channel in the high-dimensional nonlinear convolution feature, represents the high-dimensional nonlinear convolution feature after positive definiteness, represents the correction threshold, is the identity matrix, represents an orthogonal matrix; The average value of the positively definite high-dimensional nonlinear convolution features is: in, represents the high-dimensional nonlinear convolution feature with k channels after positive quantization, Represents the depthwise convolution feature matrix.

2. The method according to claim 1, characterized in that Uses Doppler-steering vector-driven simulated target signals within pre-acquired sea clutter data, including: The simulated target signal using Doppler steering vector drive in pre-acquired sea clutter data is: in, is the sea clutter data, is the Doppler steering vector, is the target signal amplitude parameter, represents the initial phase, represents the scattering coefficient, Indicates the number of target signal pulses, represents the normalized Doppler frequency.

3. The method according to claim 1, characterized in that The method of obtaining the time-frequency characteristic diagram of the echo data by using a time-frequency analysis method includes: The STFT time-frequency analysis method is used to obtain the time-frequency characteristic diagram of the echo data.

4. The method according to claim 1, wherein The number of convolutional layers of the convolutional neural network is 1 or 2.

5. The method according to claim 4, characterized in that When the number of convolutional layers of the convolutional neural network is 1, the number of network layers of the deep manifold network when the signal-to-noise ratio is greater than the first signal-to-noise ratio is 4, the number of network layers when the signal-to-noise ratio is less than the first signal-to-noise ratio and greater than the second signal-to-noise ratio is 8, and the number of network layers when the signal-to-noise ratio is less than the second signal-to-noise ratio is 6; when the number of convolutional layers of the convolutional neural network is 2, the number of network layers of the deep manifold network is 8.

6. A target detection device based on a deep convolutional feature matrix manifold network, used to implement the target detection method based on a deep convolutional feature matrix manifold network according to any one of claims 1 to 5, characterized in that: The device comprises: a data acquisition module configured to acquire echo data of simulated targets having different signal-to-noise ratios using a simulated target signal driven by a Doppler steering vector from pre-acquired sea clutter data; the Doppler steering vector being represented by a target signal amplitude parameter, the magnitude of which is related to the average magnitude of the sea clutter data, the signal-to-noise ratio, the initial phase, and the scattering coefficient; A deep convolution feature matrix construction module is used to obtain a time-frequency feature map of the echo data using a time-frequency analysis method, and to extract high-dimensional nonlinear features of the time-frequency feature map using a pre-trained convolutional neural network to obtain a deep convolution feature matrix; The target detection module is used to train a pre-built deep manifold network using the deep convolution feature matrix as a sample, so that the deep manifold network can detect targets in the echo data to be detected.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images