A radar target detection method based on shallow and deep feature fusion

Abstract

The present application relates to the field of signal processing, deep learning and radar target detection, and particularly relates to a radar target detection method based on shallow and deep feature fusion. The present application solves the problem of poor detection generalization caused by relying on shallow features in the prior art by fusing shallow multi-domain features and deep learning features. The shallow features are subjected to multi-domain analysis, weighting and normalization processing, enhancing the ability to distinguish target echoes and sea clutter. The deep features are extracted by 1D-ResNet50 and spliced with the shallow features, combining physical meaning and expression ability, and improving the feature representation ability of the model. Combined with the constant false alarm rate algorithm, the detection threshold is dynamically adjusted to ensure stability and robustness in complex marine environments. The present application significantly improves the accuracy, generalization and practicality of target detection, and has high engineering application value.

Description

Technical Field

[0001] This invention relates to the fields of signal processing, deep learning, and radar target detection. More specifically, it relates to a radar target detection method based on the fusion of shallow and deep features. This method extracts and fuses shallow and deep features of radar signals to detect targets on the sea surface. Background Technology

[0002] In recent years, sea targets have gradually become smaller and more stealthy. As the "eyes" of the maritime battlefield and maritime surveillance scenarios, radar needs to further enhance its maritime detection capabilities. A deeper understanding, more refined perception, and full utilization of the characteristics of sea clutter are crucial. Radar target detection involves discovering the presence of target echoes from complex received signals. The main approaches to target detection in cluttered environments are clutter suppression and signal enhancement. With the increasing intelligence of active radar jamming, traditional jamming suppression methods relying solely on signal processing have shown significant limitations. Furthermore, under refined observation conditions, radar echoes carry rich environmental and target information; therefore, fully utilizing this information is essential for radar target detection.

[0003] Patent CN105334507A discloses a method for detecting floating radar targets on the sea surface based on polarization multi-features. It utilizes a model decomposition method based on biunitary transformation to obtain polarization features—the energy corresponding to the dihedral scattering mechanism and the energy corresponding to the volume scattering mechanism, as well as the relative Doppler peak height features in the frequency domain. These three features are used to construct a detection decision region, and the presence of a target in the detection unit is determined based on the detector's decision region. By using the differences between sea clutter and the target in the echo signal across multiple domains as a distinguishing factor, and combining this with machine learning methods, a binary classification task is performed on the echo sequence based on these features, thereby detecting the presence of the target. With the development of deep learning, directly feeding radar echo signals into deep networks for high-dimensional feature extraction, and then detecting targets based on deep features, is gradually becoming the mainstream method. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies that only extract multi-domain differential features of sea clutter and target echoes and thus have poor detection generalization, this invention proposes a radar target detection method based on the fusion of shallow and deep features. This solves the problem of poor detection generalization caused by using only shallow features, and improves the model's representation of echo features by weighting the shallow features.

[0005] This invention mainly includes the following parts:

[0006] (I) Extraction of shallow features.

[0007] Shallow feature extraction refers to extracting the differences between sea clutter and target echoes in the time domain, frequency domain, and time-frequency domain in radar echo sequences. After multiple experiments, this invention selects six features as shallow data: relative average amplitude (RAA) in the time domain, relative Doppler peak height (RDPH) and relative Doppler vector entropy (RVE) in the frequency domain, and ridge integral (RI), number of connected regions (NR), and maximum size of connected regions (MS) in the time-frequency domain. Specifically, the relative average amplitude (RAA) in the time domain is the ratio of the average amplitude of the measured element to that of the reference element in the time domain; the relative Doppler peak height (RDPH) in the frequency domain is the ratio of the Doppler peak value of the measured element to the average Doppler peak value of the reference element; and the relative Doppler vector entropy (RVE) is the ratio of the information entropy of the measured element to that of the reference element. The three features in the time-frequency domain are obtained by performing a smooth pseudo-Wigner-Ville distribution (SPWVD) transformation on the radar signal echo, and using the discrete plane curve composed of the maxima of each time slice as the ridge integral (RI). Then, the time-frequency ridge is binarized, and the first L largest pixels of each time slice are set to 1, while other pixels are set to 0, to obtain a binary image. For any two pixels, the eight-adjacency rule is used to determine whether they are connected, thereby obtaining the number of connected regions (NR) and the maximum size of the connected regions (MS).

[0008] (ii) Weighting of shallow features.

[0009] The importance of six shallow features is ranked using the Gini index in the Classification and Regression Tree (CART) algorithm within a decision tree. The Gini index represents the uncertainty (i.e., purity) of a set. Specifically, it's calculated on different features to classify sea clutter and target echoes. A smaller Gini index indicates higher purity, meaning a lower probability of misclassifying a selected sample, making that feature the optimal choice for classification. In this invention, the Gini index is calculated for the six shallow features, and a decision tree is fitted. The features are then weighted according to their calculated importance, with a total weight of 1 for all six features.

[0010] (III) Extraction of deep features.

[0011] A one-dimensional convolutional neural network (1D-CNN) is employed for deep feature extraction from radar echo data, specifically using an improved ResNet-50 architecture, termed 1D-ResNet50. This network effectively processes time-series data and is suitable for feature learning of one-dimensional signals such as radar signals. The network first receives the raw radar echo data and encodes its features through a series of residual convolutional blocks. Each block contains multiple convolutional layers and residual connections to enhance learning capabilities and prevent gradient vanishing during training. After successive convolution and pooling operations, the network transforms the data into a 2048-dimensional feature vector. This high-dimensional feature vector 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, outputting a more compact feature representation. This step not only reduces the dimensionality of the features but also preserves key information, making the model more efficient and accurate in target classification, tracking, or other downstream tasks.

[0012] (iv) The controllability of the constant false alarm rate and the training and testing of the model.

[0013] For feature fusion, this invention operates at the feature level, concatenating deep-learned features with hand-designed features. This method balances the expressive power of deep features with the intuitiveness of shallow features. The concatenated features are passed to a linear classification layer, which maps high-dimensional feature information to a one-dimensional output. Then, a sigmoid activation layer converts this one-dimensional output into a probability value indicating the target's presence in that echo band.

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

[0015] The model was trained and tested using a large amount of radar data. During training, the model was trained to recognize target echoes from various sea surface conditions, from calm to rough. During testing, independent test sets were used to evaluate the model's generalization ability and the effectiveness of false alarm control. This ensured that the model underwent rigorous testing before actual deployment and could achieve efficient target detection under a wide range of operating conditions.

[0016] The specific technical solution adopted in this invention is as follows:

[0017] A radar target detection method based on the fusion of shallow and deep features is proposed. Addressing the target detection problem in complex sea clutter backgrounds, this method introduces an innovative detection scheme that fuses shallow multi-domain features with deep learned features. The specific steps are as follows:

[0018] S1: Data Preprocessing. The input complex-valued radar echo data is processed by absolute value and normalized to ensure the consistency of data format and the effectiveness of processing.

[0019] S2: Extracting shallow features from echo signals. Echo segments of consistent length are extracted from sea clutter and the target signal to form a dataset 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 of 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. Extracting Deep Features from the Echo Signal. The deep feature extraction part employs a 1D-ResNet50 network to learn high-level nonlinear feature representations from the echo signal. These deep features are then mapped to high-dimensional features through a linear layer for fusion with shallow features. Feature fusion uses a concatenation method at the feature level, combining the deeply learned features with weighted shallow features. This preserves the intuitive physical meaning of the shallow features while maintaining the expressive power of the deep features. The fused high-dimensional features are input into a linear classification layer, mapping them to a one-dimensional output. The Sigmoid activation function then transforms this output value into the probability of the target's existence.

[0022] S5. Model Training. To maintain a stable false alarm rate in complex marine environments, this method integrates a constant false alarm rate (CFAR) algorithm. Through testing on a pure sea clutter dataset, the detection threshold is automatically adjusted based on a set false alarm rate to ensure that a predetermined false alarm level is maintained even in complex or changing background environments. Supervised learning is then used to train the fused features, resulting in a model that effectively distinguishes target echoes from sea clutter echoes in binary classification tasks, significantly improving the radar system's target detection performance and environmental adaptability.

[0023] The beneficial effects of this invention are as follows: By fusing shallow multi-domain features with deep learning features, this invention solves the problem of poor detection generalization caused by relying on shallow features in existing technologies. Shallow features undergo multi-domain analysis, weighting, and normalization, enhancing the ability to distinguish between target echoes and sea clutter. Deep features are extracted using 1D-ResNet50 and concatenated with shallow features, possessing both physical meaning and expressive power, thus improving the model's feature representation ability. Combined with the constant false alarm rate (CFAR) algorithm, the detection threshold is dynamically adjusted to ensure stability and robustness in complex marine environments. This invention significantly improves the accuracy, generalization, and practicality of target detection, and has high engineering application value. Attached Figure Description

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

[0025] Figure 2 This is a flowchart of the radar target detection process of the present invention.

[0026] Figure 3 These are time-frequency diagrams and normalized binary diagrams comparing the target echo and sea clutter; where (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 cell; and (d) is the normalized binary diagram of the sea clutter after SPWVD time-frequency transformation.

[0027] Figure 4 The results are visualizations of the detection results for a portion of the dataset; (a), (b), (c), and (d) are the detection results on different datasets, with the lighter-colored portions representing the samples targeted for detection. Detailed Implementation

[0028] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0029] This embodiment presents a radar target detection method based on the fusion of shallow and deep features. The basic process is as follows: Figure 2 As shown, the overall network structure adopted is as follows: Figure 1 As shown.

[0030] Specifically, the following steps are included:

[0031] S1. Preprocess the complex-valued signals in the publicly available radar dataset IPIX. First, extract data from the IPIX dataset. Each data set consists of a complex-valued matrix of size 131072×14, where 131072 represents the slow-time dimension sampling points (approximately 131 seconds), and 14 represents the fast-time dimension range cells. Each data set consists of one primary target echo cell and several secondary target echo cells and sea clutter cells. In this invention, only the primary target echo cell is selected as the target cell. The target cell and sea clutter cells are each truncated into 512-length echo segments as the dataset.

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

[0033] For the time-domain feature-relative average amplitude (RAA), after calculating the amplitude values ​​of the complex-valued data, the sea clutter element is used as the reference element, and the ratio of the average amplitude of each measured element to the average amplitude of the reference element, RAA(x,x), is calculated. k ), defined as follows:

[0034]

[0035] Where x is the echo of the unit under test, x k (k = 1, 2, ..., K) is the echo of the kth reference unit. It is the average amplitude of the echo. When the echo contains a target, the RAA value is relatively large, ranging from tens to hundreds; otherwise, it is relatively small.

[0036] For the frequency domain features—relative Doppler peak height (RDPH) and relative Doppler vector entropy (RVE)—after calculating the Doppler values ​​by fast Fourier transform of the complex-valued data in the slow time dimension, the sea clutter cell is used as the reference cell. The ratio of the Doppler peak value of each measured cell to the average Doppler peak value of the reference cell, RDPH(z,z), is calculated. 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 RVE is the Doppler amplitude spectrum of the reference cell, DPH is the Doppler peak height of the echo cell, and VE is the information entropy of the echo cell. When the echo contains a target, RPH takes a large positive value, ranging from tens to hundreds. For sea clutter, RPH takes a positive value of around 1. When the echo contains a target, RVE takes a small value less than 1. For sea clutter, RVE is tens of times larger than the target RVE.

[0039] For the time-frequency domain features—ridge integral (RI), number of connected components (NR), and maximum size of connected components (MS)—after performing a smooth pseudo-Wigner-Ville distribution (SPWVD) transformation on each complex-valued data, the discrete plane curve composed of the maxima of each time slice is taken as the ridge integral (RI). The time-frequency ridge is then binarized, with the first 5 largest pixels of each time slice set to 1 and the other pixels set to 0, resulting in a binary image. For any two pixels, the eight-adjacency rule is used to determine whether they are connected, thus obtaining the number of connected components (NR) and the maximum size of connected components (MS), defined as follows:

[0040]

[0041]

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

[0043] For echoes containing targets, the RI value is relatively large, ranging from tens to hundreds, while for sea clutter, the RI value is relatively small, ranging from tens to hundreds.

[0044] Both NR and MS are geometric features extracted from NTFD. First, the top L largest pixels at each time point in the NTFD are extracted, and these top L largest pixels are set to 1, while other pixels are set to 0. This generates the threshold NTFD for the binary image (data in this embodiment is as follows). Figure 3 (As shown). Next, all pixels with a value of 1 in NTFD are connected into regions using the eight-neighbor standard. The number of connected regions is the NR value, and the number of pixels contained in the largest connected region is the MS value. 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, the specific values ​​depending on the pulse accumulation number of the echo cell.

[0045] S3. Weighting the shallow features of the echo signal. Based on the six features extracted in S2, CART is used for classification, and the Gini index is calculated to determine the importance of each feature. The six features are then weighted with a total weight of 1. The Gini index (D,a) is calculated as follows:

[0046]

[0047] Where v is the v-th subset, D is the total dataset, and pkv Let be the probability that a sample belongs to class k in the v subset, Gini(D,a) be the Gini index for classification using feature a, and Gini(D) be the Gini index for all features.

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

[0049] S4. Extracting Deep Features from Echo Signals. A 1D-ResNet50 algorithm is used to extract deep features from the radar echo data. First, the modulus of the complex-valued radar echo data is calculated and used as input. Then, features are encoded through a series of residual convolutional blocks. Each block contains multiple convolutional layers and residual connections to enhance learning ability and prevent gradient vanishing during training, transforming the 512-dimensional data 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, outputting a more compact feature representation.

[0050] S5. Train the model, obtain a threshold using a controllable constant false alarm rate (CFAR), and test it. Divide the dataset obtained in S1 into training and test sets proportionally (e.g., a 7:3 ratio). Extract shallow features from the training set and fuse them with the deep features obtained through the network. Train the model using known labels through supervised learning, employing the binary cross-entropy loss function (BCELoss). After model training and fitting, test it on a pure sea clutter dataset. Automatically adjust the detection threshold according to the set CFAR algorithm, and then use this threshold for classification on the test set to statistically analyze the detection results. Visualization results of some data are shown below. Figure 4 As shown, the light-colored areas are the samples that are detected as targets, and the distance cells with larger light-colored areas are the target cells. It can be seen that the method proposed in this invention detects most of the target samples in the target cells 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; the input complex-valued radar echo data is subjected to absolute value processing and normalization to ensure the consistency of data format and the effectiveness of processing; S2: extracting the shallow features of the echo signal; the length-consistent echo segment is intercepted from the sea clutter and target signal as a data set for shallow and deep feature extraction; in the shallow feature extraction stage, a multi-domain analysis method is adopted to extract six key features from the time domain, frequency domain and time-frequency domain, including time domain feature-relative average amplitude RAA, frequency domain feature-relative Doppler peak height RDPH and relative Doppler vector entropy RVE, and time-frequency domain feature-ridge integral RI, connected region number NR and connected region maximum size MS; S3. Weighting the shallow features of the echo signal; the shallow features are subjected to Gini index weighting and normalization processing to further optimize their representativeness in target detection; S4. Extracting the deep features of the echo signal; The deep feature extraction part adopts a 1D-ResNet50 network to learn high-level nonlinear feature representation from the echo signal, and maps the deep features to high-dimensional features through a linear layer for fusion with the weighted shallow features; Feature fusion adopts a splicing method at the feature level to combine the deep learning features with the weighted shallow features; the fused high-dimensional features are input into a linear classification layer to be mapped to one-dimensional output, and the output value is converted into the probability of target existence through a Sigmoid activation function; S5. Training the model; integrating the constant false alarm rate CFAR algorithm, through testing on the pure sea clutter data set, the detection threshold is automatically adjusted according to the set false alarm rate to ensure that the predetermined false alarm level is maintained in complex or changing background environment; and the fused features are trained through supervised learning.

2. The 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 value of the complex value data, the sea clutter cell is taken as the reference cell, and the ratio of the average amplitude of each to-be-tested cell and the reference cell is calculated RAA(x, x k ), which is defined as follows: where x is the unit echo to be measured, x k (k = 1, 2, …, K) is the kth reference unit echo, is the average amplitude of the echoes.

3. The radar target detection method based on shallow and deep feature fusion according to claim 1, characterized in that, For the frequency domain features, the relative Doppler peak height RDPH and the relative Doppler vector entropy RVE, the complex data is first transformed into the Doppler domain by taking the fast Fourier transform (FFT) in the slow time dimension. Then, the ratio of the Doppler peak value of each cell under test to the average Doppler peak value of the reference cells RDPH(z,z k ) and the ratio of the information entropy of the cell under test to the information entropy of the reference cells RVE(z,z k ) are calculated, which are defined as follows: where z is the Doppler amplitude spectrum of the unit to be tested, z k is the Doppler amplitude spectrum of the reference unit, DPH is the Doppler peak height of the return unit, and VE is the information entropy of the return unit.

4. The radar target detection method based on shallow and deep feature fusion according to claim 1, characterized in that, For time-frequency domain features-ridge integral RI, connected region number NR and connected region maximum size MS, after smoothing pseudo Wigner-Ville distribution transformation of each complex-valued data, the discrete planar curve composed of the maximum value of each time slice is taken as the ridge integral RI, the time-frequency ridge is subjected to binaryzation transformation, 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; according to the eight-neighborhood rule, it is judged whether any two pixels are connected to obtain the connected region number NR and the connected region maximum size MS, which are defined as follows: where n is the total number of back-echo units, l is the length of the back-echo unit, 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 transform, and RI(z, z k ) is the ridge integral in the time-frequency diagram. Both NR and MS are geometric features extracted from NTFD; first, the first L maximum pixel points at each time point on NTFD are extracted, and the first L maximum pixel points are taken as 1 and the other pixel points are taken as 0 to produce a threshold NTFD of the binary image; next, all the pixels with a value of 1 in NTFD form a connected region through the eight-neighborhood standard; then, the number of connected regions is the value of NR, and the number of pixel points contained in the largest connected region 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: using constant false alarm rate CART to classify according to the 6 features extracted in S2, calculating the Gini index to obtain the importance degree between features so as to weight the 6 features, and the total weight is 1; the Gini index Gini(D,a) is calculated as follows: where v is the vth subset, D is the total dataset, p kv is the probability of a sample belonging to the kth class being in the vth subset, Gini(D, a) is the Gini index for classification using feature a, and Gini(D) is the Gini index for all features.

Images