Radar target detection method based on shallow and deep feature fusion

Abstract

The invention relates to the field of signal processing, deep learning and radar target detection, in particular to a radar target detection method based on shallow and deep feature fusion. By fusing the shallow multi-domain features and the deep learning features, the problem of poor detection generalization caused by dependence on shallow features in the prior art is solved. The shallow features are subjected to multi-domain analysis, weighting and normalization processing, so that the distinguishing capability of target echoes and sea clutters is enhanced; the deep layer features are extracted through 1D-ResNet50 and spliced with the shallow layer features, the physical significance and the expression ability are achieved, and the feature representation ability of the model is improved. And in combination with a constant false alarm rate algorithm, a detection threshold value is dynamically adjusted, and the stability and robustness in a complex marine environment are ensured. According to the method, the precision, generalization and practicability of target detection are remarkably improved, and the method has high engineering application value.

Description

Technical Field

[0001] The present invention relates to the fields of signal processing, deep learning and radar target detection, and more specifically, to a radar target detection method based on the fusion of shallow and deep features, which is a method for respectively extracting and fusing shallow features and deep features of radar signals to detect sea surface targets. Background Art

[0002] In recent years, sea targets have gradually become "miniaturized" and "stealthier". Radar, as the "eyes" of the sea battlefield and sea detection scenes, is crucial to further enhance the radar's sea detection capabilities and to deeply understand, finely perceive and fully utilize the characteristics of sea clutter. Radar target detection is to discover the existence of target echoes from complex received signals. The main way to solve target detection under clutter background is clutter suppression and signal enhancement. With the continuous intelligence of radar active jammers, traditional jammer suppression methods that rely solely on signal processing methods have shown obvious shortcomings. At the same time, under refined observation conditions, radar echoes carry rich environmental and target information, so making full use of this information is a necessary means for radar target detection.

[0003] Patent CN105334507A discloses a method for detecting floating radar targets on the sea surface based on polarization multi-features. The polarization feature - the energy corresponding to the dihedral angle scattering mechanism and the energy corresponding to the volume scattering mechanism, as well as the relative Doppler peak height feature in the frequency domain are obtained by using a model decomposition method based on double unitary changes. The decision area of ​​the detection is constructed by these three features, and whether there is a target in the detection unit is judged according to the decision area of ​​the detector. By using the differential features of sea clutter and targets in multiple domains in the echo signal as the basis for distinction, and combining the machine learning method to perform a binary classification task of the echo sequence based on this series of features, it is detected whether the target exists. With the development of deep learning, sending the radar echo signal directly into the deep network for high-dimensional feature extraction, and thus performing target detection based on deep features has gradually become the mainstream method. Summary of the invention

[0004] In order to overcome the defects of the prior art that only multi-domain difference features of sea clutter and target echoes are extracted and classified, resulting in poor detection generalization, the present invention proposes a radar target detection method based on the fusion of shallow and deep features, which solves the problem of poor detection generalization caused by only using shallow features, and improves the representation of the model for echo features by weighting the shallow features.

[0005] The present invention mainly includes the following parts:

[0006] (A) Extraction of shallow features.

[0007] Shallow feature extraction refers to extracting the differential features of sea clutter and target echoes in the radar echo sequence in the time domain, frequency domain and time-frequency domain. After many experiments, the present invention selects the relative average amplitude (RAA) in the time domain, the relative Doppler peak height (RDPH) and the relative Doppler vector entropy (RVE) in the frequency domain, and the ridge integral (RI) in the time-frequency domain, the number of connected regions (NR) and the maximum size of the connected regions (MS) as shallow data. Among them, the relative average amplitude (RAA) in the time domain is the average amplitude ratio of the unit to be tested to the reference unit in the time domain, the relative Doppler peak height (RDPH) in the frequency domain refers to the ratio of the Doppler peak of the unit to be tested to the average Doppler peak of the reference unit, and the relative Doppler vector entropy (RVE) refers to the ratio of the information entropy of the unit to be tested to the information entropy of the reference unit. The three features in the time-frequency domain are: after the radar signal echo is transformed into a smoothed pseudo-Wigner-Ville distribution (SPWVD), the discrete plane curve composed of the maximum values ​​of each time slice is used as the ridge integral (RI), and then the time-frequency ridge is binarized, the first L maximum pixels of each time slice are taken as 1, and the other pixels are taken as 0 to obtain a binary image; for any two pixels, the eight-adjacent rule is used to determine whether they are connected, so as to obtain the number of connected regions (NR) and the maximum size of the connected region (MS).

[0008] (ii) Weighting of shallow features.

[0009] The importance of 6 shallow features is sorted by the Gini index in the classification and regression tree (CART) algorithm in the decision tree. The Gini index represents the uncertainty (i.e., purity) of a set, that is, the Gini index for classifying sea clutter and target echoes is calculated on different features. The smaller the Gini index, the higher the purity of the set, and the smaller the probability that the selected sample in the set is misclassified, then this feature can be classified as the optimal feature. In the present invention, the Gini index is calculated for 6 shallow features and a decision tree is fitted, and the corresponding features are weighted according to the calculated importance, wherein the total weight of the 6 features is 1.

[0010] (III) Extraction of deep features.

[0011] A one-dimensional convolutional neural network (1D-CNN) is used to extract deep features from radar echo data. Specifically, an improved ResNet-50 architecture, called 1D-ResNet50, is used. This network can effectively process time series data and is suitable for feature learning of one-dimensional signals such as radar signals. The network first receives the original radar echo data and encodes the features through a series of residual convolution blocks. Each block contains multiple convolution layers and residual connections to enhance learning ability and prevent gradient disappearance during training. After continuous convolution and pooling operations, the network converts the data into a 2048-dimensional feature vector. This high-dimensional feature contains rich spatial and frequency information extracted from the original radar data, providing strong data support for subsequent signal processing and target recognition. Subsequently, these features are passed to a linear transformation layer, which maps the 2048-dimensional features to 128 dimensions for outputting a more compact feature representation. This step not only reduces the dimension of the features, but also retains key information, making the model more efficient and accurate when performing target classification, tracking or other downstream tasks.

[0012] (IV) Controllability of constant false alarm rate and training and testing of models.

[0013] For feature fusion, the present invention chooses to operate at the feature level and chooses to splice deep learning features with manually designed features. This method can take into account the expressiveness of deep features and the intuitiveness of shallow features. The spliced ​​features are passed to a linear classification layer, which is responsible for mapping high-dimensional feature information to one-dimensional output. Next, through a sigmoid activation layer, the output of this dimension is converted into the probability value of the target existing in the echo band.

[0014] In terms of false alarm rate control, a constant false alarm rate (CFAR) algorithm is integrated into the algorithm used in the present invention to ensure that a predetermined false alarm level can be maintained under different marine environments. By testing on a pure sea clutter data set, the detection threshold can be automatically adjusted according to the set false alarm rate. Adjusting the threshold in this way allows a stable false alarm rate to be maintained in practical applications even in complex or changing clutter environments.

[0015] The training and testing of the model is carried out by using a large amount of radar data. In the training phase, the model is trained to identify target echoes from different types of sea surface conditions, from calm sea surface to rough sea state. In the testing phase, an independent test set is used to evaluate the generalization ability of the model and the effect of false alarm control. This ensures that the model has been rigorously tested before actual deployment and can achieve efficient target detection under a wide range of operating conditions.

[0016] The technical solution adopted by the present invention is specifically as follows:

[0017] A radar target detection method based on the fusion of shallow and deep features is proposed to solve the problem of target detection in complex sea clutter background by fusing shallow multi-domain features with deep learning features. The specific steps are as follows:

[0018] S1: Data preprocessing: Perform absolute value processing and normalization on the input complex-valued radar echo data to ensure the consistency of the data format and the effectiveness of the processing.

[0019] S2: Extract shallow features of echo signals. Echo segments of the same length are cut from sea clutter and target signals as data sets for shallow and deep feature extraction. In the shallow feature extraction stage, a multi-domain analysis method is used to extract six key features from the time domain, frequency domain and time-frequency domain, including time domain features - relative average amplitude (RAA), frequency domain features - relative Doppler peak height (RDPH) and relative Doppler vector entropy (RVE), and time-frequency domain features - ridge integral (RI), number of connected regions (NR) and maximum size of connected regions (MS).

[0020] S3. Weighting the shallow features of the echo signal. These shallow features are weighted and normalized by the Gini index to further optimize their representativeness in target detection.

[0021] S4. Extract deep features of echo signals. The deep feature extraction part uses the 1D-ResNet50 network to learn high-level nonlinear feature representations from echo signals, and maps deep features to high-dimensional features through linear layers for fusion with shallow features. Feature fusion uses a splicing method at the feature level to combine deep learning features with weighted shallow features, which not only retains the intuitive physical meaning of shallow features, but also takes into account the expressive power of deep features. The fused high-dimensional features are input into a linear classification layer, mapped to a one-dimensional output, and the output value is converted into the probability of the target's existence through the Sigmoid activation function.

[0022] S5. Train the model. In order to maintain a stable false alarm rate in a complex ocean environment, this method integrates a constant false alarm rate (CFAR) algorithm. Through testing on a pure sea clutter dataset, the detection threshold is automatically adjusted according to the set false alarm rate to ensure that the predetermined false alarm level can be maintained in a complex or changing background environment. The fused features are trained through supervised learning, and the resulting model can effectively distinguish target echoes from sea clutter echoes in a binary classification task, significantly improving the target detection performance and environmental adaptability of the radar system.

[0023] Beneficial effects of the present invention: The present invention solves the problem of poor detection generalization caused by reliance on shallow features in the prior art by fusing shallow multi-domain features with deep learning features. Shallow features are subjected to multi-domain analysis, weighting and normalization processing to enhance the ability to distinguish between target echoes and sea clutter; deep features are extracted through 1D-ResNet50 and spliced ​​with shallow features, which have both physical meaning and expression ability, and improve the feature characterization ability of the model. Combined with the constant false alarm rate (CFAR) algorithm, the detection threshold is dynamically adjusted to ensure stability and robustness in complex marine environments. The present invention significantly improves the accuracy, generalization and practicality of target detection, and has high engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the overall structure of the network model of radar target detection of the present invention;

[0025] Figure 2 It is a radar target detection flow chart of the present invention.

[0026] Figure 3 It is the time-frequency diagram and normalized binary diagram of the comparison between the target echo and the sea clutter; among them, (a) is the time-frequency diagram of the target echo; (b) is the normalized binary diagram of the target echo after SPWVD time-frequency transformation; (c) is the time-frequency diagram of the sea clutter unit; (d) is the normalized binary diagram of the sea clutter after SPWVD time-frequency transformation.

[0027] Figure 4 It is the visual detection results of some data sets; (a), (b), (c), and (d) are the detection results on different data respectively, and the light-colored part is the sample detected as the target. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0029] In this embodiment, a radar target detection method based on shallow and deep feature fusion is described. The basic process is as follows: Figure 2 As shown in Figure 2, the overall network structure used is as follows: Figure 1 shown.

[0030] The specific steps include:

[0031] S1. Preprocess the complex-valued signals in the public radar data set IPIX. First, extract data from the IPIX data set. Each set of data consists of a complex-valued matrix of size 131072×14, where 131072 is a slow time dimension sampling point, about 131 seconds, and 14 is a fast time dimension distance unit. Each set of data consists of a main target echo unit and several secondary target echo units and sea clutter units. In the present invention, only the main target echo unit is selected as the target unit, and the target unit and the sea clutter unit are respectively cut into echo segments of length 512 as the data set;

[0032] S2. Extract the shallow features of the echo signal. Calculate the corresponding features in the time domain, frequency domain, and time-frequency domain for each complex-valued data in the data set.

[0033] For the time domain feature - relative average amplitude (RAA), after calculating the amplitude of the complex-valued data, the sea clutter unit is used as the reference unit to calculate the ratio of the average amplitude of each unit under test and the reference unit RAA (x, x k ), defined as follows:

[0034]

[0035] Among them, x is the echo of the unit under test, x k (k=1,2,…,K) is the kth reference unit echo, It is the average amplitude of the echo. When the echo contains a target, the RAA value is larger, ranging from about ten to several dozen, otherwise it is smaller.

[0036] For the frequency domain features - relative Doppler peak height (RDPH) and relative Doppler vector entropy (RVE), the Doppler value is obtained by fast Fourier transform of the complex-valued data in the slow time dimension. Then, the sea clutter unit is used as the reference unit to calculate the ratio of the Doppler peak value of each unit to be tested to the average Doppler peak value of the reference unit RDPH (z, z k ), and the ratio of the information entropy of the unit under test to the information entropy of the reference unit RVE(z,z k ), defined as follows:

[0037]

[0038] Where z is the Doppler amplitude spectrum of the unit under test, z k is the Doppler amplitude spectrum of the reference unit, DPH is the Doppler peak height of the echo unit, and VE is the information entropy of the echo unit. When the echo contains a target, RDPH takes a larger positive value ranging from a dozen to dozens. For sea clutter, RDPH takes a positive value of about 1; when the echo contains a target, RVE takes a smaller value less than 1. For sea clutter, RVE is dozens of times greater than the target RVE.

[0039] For the time-frequency domain features - ridge integral (RI), number of connected regions (NR) and maximum size of connected regions (MS), after smooth pseudo Wigner-Ville distribution (SPWVD) transformation is performed on each complex-valued data, the discrete plane curve composed of the maximum value of each time slice is used as the ridge integral (RI), the time-frequency ridge is binarized, the first 5 maximum pixels of each time slice are taken as 1, and the other pixels are taken as 0 to obtain a binary image; for any two pixels, according to the eight-adjacent rule, it is judged whether they are connected, so as to obtain the number of connected regions (NR) and the maximum size of connected regions (MS), which are defined as follows:

[0040]

[0041]

[0042] Where n is the total number of echo units, l is the length of the echo unit, and NTFD(n,l|z,z k ) is the normalized time-frequency diagram, SPWVD(n,l|z) is the time-frequency diagram after SPWVD time-frequency transformation, RI(z,z k ) is the ridge integral in the time-frequency plot.

[0043] For echoes containing targets, the RI value is relatively large, approximately tens to hundreds. For sea clutter, the RI value is relatively small, approximately a dozen to dozens.

[0044] NR and MS are both geometric features extracted from NTFD. First, extract the first L maximum pixel points at each time point on NTFD, set the first L maximum pixel points to 1, and the other pixel points to 0, and then generate the threshold NTFD of the binary image (the data in this embodiment is as follows Figure 3 As shown). Next, all pixels with a value of 1 in NTFD form a connected area through the eight-neighborhood standard. Then, the number of connected areas is the value of NR, and the number of pixels contained in the largest connected area is the value of MS. When the radar echo contains a target, it has a smaller NR value and a larger MS value; conversely, sea clutter has a larger NR value and a smaller MS value, and the specific value depends on the number of pulse accumulations of the echo unit.

[0045] S3. Weight the shallow features of the echo signal. Use CART to classify the 6 features extracted in S2, calculate the Gini index, and get the importance of the features to weight the 6 features, with a total weight of 1; the Gini index Gini (D, a) is calculated as follows:

[0046]

[0047] Among them, v is the vth subset, D is the total data set, and pkv is the probability of a sample belonging to the kth class in the vth subset, Gini(D,a) is the Gini index for classification using feature a, and Gini(D) is the Gini index of all features.

[0048] By calculating the difference in the Gini index before and after using feature a for classification, we can obtain ΔGini(D,a)=Gini(D)-Gini(D,a). The larger ΔGini(D,a) is, the greater the contribution of feature a is, and therefore the more important feature a is.

[0049] S4. Extract deep features of echo signals. 1D-ResNet50 is used to extract deep features from radar echo data. First, the modulus of complex-valued radar echo data is calculated as input data, and then feature encoding is performed through a series of residual convolution blocks. Each block contains multiple convolution layers and residual connections to enhance learning ability and prevent gradient disappearance during training. The 512-dimensional data is converted into a 2048-dimensional feature vector. Subsequently, these features are passed to a linear transformation layer, which maps the 2048-dimensional features to 128 dimensions and outputs a more compact feature representation.

[0050] S5. Train the model, obtain the threshold through the controllable constant false alarm rate and test it. Divide the data set obtained in S1 into training set and test set in proportion (it can be divided in a ratio of 7:3), extract the shallow features of the training set and the deep features obtained through the network, and fuse them. Use supervised learning to train the model with known labels, and the loss function uses the binary cross entropy loss function (BCELoss). After the model is trained and fitted, it is tested on a pure sea clutter data set. The detection threshold is automatically adjusted according to the set constant false alarm rate algorithm, and then this threshold is used in the classification of the test set to perform statistics on the detection results. The detection visualization results of some data are shown in the figure. Figure 4 As shown, the light-colored part is the sample detected as the target, wherein the distance unit with a larger light-colored area is the target unit. It can be seen that the method proposed in the present invention detects most of the target samples in the target unit and has effective detection performance.

Claims

1. A radar target detection method based on shallow and deep feature fusion, characterized in that: The specific steps are as follows: S1: Data preprocessing: Perform absolute value processing and normalization on the input complex-valued radar echo data to ensure the consistency of data format and the effectiveness of processing; S2: Extract shallow features of echo signals; intercept echo segments of the same length from sea clutter and target signals as data sets for shallow and deep feature extraction; in the shallow feature extraction stage, a multi-domain analysis method is used to extract six key features from the time domain, frequency domain and time-frequency domain, including time domain features - relative average amplitude RAA, frequency domain features - relative Doppler peak height RDPH and relative Doppler vector entropy RVE, and time-frequency domain features - ridge integral RI, number of connected regions NR and maximum size of connected regions MS; S3. weighting the shallow features of the echo signal; the shallow features are weighted and normalized by the Gini index to further optimize their representativeness in target detection; S4. Extracting the depth feature of the echo signal; The deep feature extraction part uses the 1D-ResNet50 network to learn high-level nonlinear feature representations from the echo signal, and maps the deep features into high-dimensional features through the linear layer for fusion with the shallow features; Feature fusion uses a splicing method at the feature level to combine deep learning features with weighted shallow features. The fused high-dimensional features are input into a linear classification layer, mapped into a one-dimensional output, and the output value is converted into the probability of the target existence through the Sigmoid activation function. S5. Train the model; integrate the constant false alarm rate CFAR algorithm, and automatically adjust the detection threshold according to the set false alarm rate through testing on a pure sea clutter dataset to ensure that the predetermined false alarm level can be maintained in complex or changing background environments; and train the fused features through supervised learning.

2. A radar target detection method based on shallow and deep feature fusion according to claim 1, characterized in that: For the time domain feature - relative average amplitude RAA, after calculating the amplitude of the complex-valued data, the sea clutter unit is used as the reference unit to calculate the ratio of the average amplitude of each unit to be tested and the reference unit RAA(x,x k ), defined as follows: Among them, x is the echo of the unit under test, x k (k=1,2,…,K) is the kth reference unit echo, is the average amplitude of the echo.

3. The radar target detection method based on shallow and deep feature fusion according to claim 1 is characterized in that: For the frequency domain features - relative Doppler peak height RDPH and relative Doppler vector entropy RVE, the Doppler value is obtained by fast Fourier transform of the complex valued data in the slow time dimension. Then, the sea clutter unit is used as the reference unit to calculate the ratio of the Doppler peak value of each unit to be tested to the average Doppler peak value of the reference unit RDPH(z,z k ), and the ratio of the information entropy of the unit under test to the information entropy of the reference unit RVE(z,z k ), defined as follows: Where z is the Doppler amplitude spectrum of the unit under test, z k is the Doppler amplitude spectrum of the reference unit, DPH is the Doppler peak height of the echo unit, and VE is the information entropy of the echo unit.

4. The radar target detection method based on shallow and deep feature fusion according to claim 1, characterized in that: For the time-frequency domain features - ridge integral RI, number of connected regions NR and maximum size of connected regions MS, after smoothing pseudo-Wigner-Ville distribution transformation for each complex-valued data, the discrete plane curve composed of the maximum value of each time slice is used as the ridge integral RI, and the time-frequency ridge is binarized, the first 5 maximum pixels of each time slice are taken as 1, and the other pixels are taken as 0 to obtain a binary image; for any two pixels, according to the eight-adjacent rule, it is judged whether they are connected, so as to obtain the number of connected regions NR and the maximum size of connected regions MS, which are defined as follows: Where n is the total number of echo units, l is the length of the echo unit, and NTFD(n,l|z,z k ) is the normalized time-frequency diagram, SPWVD(n,l|z) is the time-frequency diagram after SPWVD time-frequency transformation, RI(z,z k ) is the ridge integral in the time-frequency diagram; Both NR and MS are geometric features extracted from NTFD. First, extract the first L largest pixels at each time point on NTFD, set the first L largest pixels to 1, and set the other pixels to 0, to generate the threshold NTFD of the binary image. Next, all pixels with a value of 1 in NTFD form a connected area through the eight-neighborhood standard. Then, the number of connected areas is the value of NR, and the number of pixels contained in the largest connected area is the value of MS.

5. The radar target detection method based on shallow and deep feature fusion according to claim 1, characterized in that: S3 is as follows: According to the 6 features extracted from S2, the constant false alarm rate CART is used for classification, and the Gini index is calculated to obtain the importance of the features and weight the 6 features, with a total weight of 1; the Gini index Gini (D, a) is calculated as follows: Among them, v is the vth subset, D is the total data set, and p kv is the probability of a sample belonging to the kth class in the vth subset, Gini(D,a) is the Gini index for classification using feature a, and Gini(D) is the Gini index of all features.

Images