A method for detecting weak targets on the sea surface based on multimodal time-frequency image fusion

Abstract

This invention discloses a method for detecting weak targets on the sea surface based on multimodal time-frequency image fusion, belonging to the field of radar signal processing technology. It converts a one-dimensional signal into three grayscale images with complementary characteristics using three signal processing methods: normalized wavelet transform, normalized smoothed pseudo-Wigner-Vell transform, and recursive graph. The resulting two-dimensional grayscale images are then mapped to the R, G, and B channels using weighted coefficients for feature fusion, generating a fused color image. Subsequently, an improved multimodal EfficientNetV2s target detection framework is constructed. By optimizing the decision threshold of the Softmax layer, weak signal detection with a controllable false alarm rate is achieved. This invention overcomes the limitations of existing time-domain and frequency-domain methods by integrating three time-frequency transformation methods. Furthermore, this method constructs an improved EfficientNetV2s network architecture detection method, designs a loss function to address the imbalanced sample problem, and incorporates an attention mechanism to enhance the model's sensitivity to key information, thereby improving detection accuracy under low sea state conditions.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, specifically to a method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion. Background Technology

[0002] Sea clutter refers to the clutter signals reflected from the sea surface in radar systems. These signals interfere with the detection of objects on the sea surface (such as ships, small boats, or buoys), reducing the performance of the radar system. Therefore, researching methods for detecting weak signals in sea clutter backgrounds is of great significance. Numerous research methods have been proposed to address the problem of weak signal detection in sea clutter backgrounds. Hu et al. proposed a fractal detector using the Hurst exponent as the test statistic, while Shui et al. proposed a detector based on three features combined with a time-frequency domain detector, which greatly improved detection performance. In subsequent research, more and more feature detection methods have been proposed, such as Suo et al.'s PCA-based detector based on eight features, and Zhao et al.'s four-feature detection method improved using the FAST algorithm. With the continuous advancement of deep learning technology, some recurrent neural networks have been widely used in sea clutter research. Yan et al. proposed a detection method based on LSTM error frequency domain transformation, which detects weak signals in the frequency domain by calculating the Doppler spectrum of the error, improving the weak signal detection performance of the LSTM model. De Jong et al. proposed a CNN-LSTM hybrid architecture monitoring model that uses time-varying amplitude fluctuations to distinguish targets. However, traditional feature detection requires feature searching and analysis, leading to missed detections and feature redundancy; while methods that directly use one-dimensional time series cannot effectively utilize signal features. Considering the complexity of data and practical needs, detection methods combining image and deep learning have emerged as a viable option. Zhao et al. chose MTF to preprocess fault sequences and input them into a CNN for classification, achieving good results. Liu used RP and GADF to process sea clutter signals, enhancing the acquisition of key information. Zhang et al. used STFT to convert one-dimensional signals into time-frequency images and then used neural networks for classification, improving accuracy.

[0003] However, time-frequency maps obtained by methods such as STFT can only represent time series from the frequency domain perspective, and cannot represent the shape differences of time series in the time-space domain. Methods based on GADF, MTF, and RP to convert time series into images can only represent the shape of the time series in the time domain, and cannot show distinguishing features in the frequency domain. Furthermore, current time series similarity measurement methods based on images and CNNs can only determine similarity based on image information obtained from a single imaging mechanism, and cannot simultaneously consider information in the time and frequency domains. Moreover, existing algorithms for detecting small targets on the sea surface perform poorly in low sea states. Therefore, a method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion is needed to address these problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion, so as to solve the problems existing in the prior art mentioned in the background art above.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion includes the following steps:

[0007] S1: Obtain the signal to be detected, which includes sea clutter signal and target echo signal. Divide the signal to be detected into training data and test data. Perform sliding window processing on the obtained data to obtain several signals to be detected.

[0008] S2: For each segment of the signal to be measured, normalized smooth pseudo-Wigner transform, normalized continuous wavelet transform and recursive graph are applied to obtain three two-dimensional grayscale images with complementary properties.

[0009] S3: The obtained two-dimensional grayscale image is mapped to the R, G, and B channels through weighted coefficients for feature fusion to generate a fused color image;

[0010] S4: Construct an improved classification framework for EfficientNetV2s, input the obtained images into the classifier, and detect weak signals.

[0011] Preferably, in step S1, the ratio of training data to test data is 7:3, the width of the sliding window is 1024, the delay is 64, and the length of the signal to be tested is 1024.

[0012] Preferably, in S2, the smoothed pseudo-Wigner transform SPWVD, for a given complex signal z = [z(1), z(2), ..., z(N)] T Its SPWVD formula is as follows:

[0013]

[0014] In the formula, h(τ) is the frequency smoothing window, and g(st) is the time smoothing window. Using the clutter cell of the sea clutter as the reference cell, the mean μ and variance σ of its SPWVD are calculated. The mean and variance are then used to normalize each cell under test to obtain the NSPWVD, as shown in the following formula:

[0015]

[0016] When detecting weak targets in sea clutter, the target echo can be modeled as a piecewise linear frequency modulated (LFM) signal. NSPWVD has a better energy focusing effect on piecewise LFM signals than time-frequency analysis methods such as STFT.

[0017] Preferably, in step S2, the mathematical definition of the continuous wavelet transform (CWT) is as follows:

[0018]

[0019] In the formula, α is the scale parameter, τ is the translation parameter, and ψ * (t) is the complex conjugate of the mother wavelet function. Morlet wavelet is selected as the mother wavelet, and clutter cells of sea clutter are used as reference cells. The mean μ and variance σ of its CWT are calculated. This mean and variance are used to normalize each cell to enhance the time Doppler characteristics of the sea clutter target echo, and the NCWT is obtained. The formula is as follows:

[0020]

[0021] NCWT has higher frequency resolution for low-frequency components and higher time resolution for high-frequency components, and can effectively capture the local time-frequency characteristics of sea clutter signals.

[0022] Preferably, in step S2, the recursive graph process is as follows:

[0023] Step 1: Select an appropriate embedding dimension m and time delay τ, and reconstruct the phase space of the one-dimensional time series to obtain the following spatial vector:

[0024] X i ={x i ,x i+ τ,…,x i+(m-1)τ}

[0025] Step 2: Calculate the Euclidean distance between each pair of trajectories after phase space reconstruction:

[0026] d ij =‖X i -X j ||

[0027] Step 3: Construct the recursive matrix, the specific equation is as follows:

[0028] R ij =Heaviside(ε-d ij )

[0029]

[0030] In the formula, ε is the distance threshold, which determines the similarity criterion, and Heaviside is the step function;

[0031] Step 4: Generate a visual RP. By applying grayscale mapping, different distance values ​​in the distance matrix are mapped to different grayscale values, thereby obtaining a grayscale image that encapsulates the complete features of the time series and forming a texture map that intuitively displays the dynamic characteristics of the time series.

[0032] Preferably, in step S4, the improved EfficientNetV2s method uses EfficientNetV2s as the backbone classification network, introduces an attention mechanism, dynamically enhances the feature responses of important channels through convolutional layers and SE layers, introduces Focal Loss to replace cross-entropy loss, reduces the loss contribution of sea clutter by adjusting the focusing parameter γ and class weight α, and enhances the model's attention to weak targets. The formula is as follows:

[0033] FL(pt)=-α(1-p t ) γ log(p t )

[0034] In the formula, p t This represents the model's predicted probability for the true class.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention integrates three time-frequency conversion methods, overcoming the limitations of existing time-domain and frequency-domain methods, and improving detection accuracy and robustness. Furthermore, this method constructs an improved detection method based on the EfficientNetV2s network architecture, designs a loss function to address the imbalanced sample problem, and incorporates an attention mechanism to enhance the model's sensitivity to key information, thereby improving detection accuracy under low sea state conditions. The measured dataset includes different sea areas and sea states, effectively promoting the application of small target detection technology on the sea surface in a wider range of scenarios. Attached Figure Description

[0037] Figure 1 This is a flowchart of the detection process of the present invention.

[0038] Figure 2 This is a structural diagram of the EfficientNetV2s model of the present invention.

[0039] Figure 3 This is a diagram illustrating the false alarm control strategy of the present invention.

[0040] Figure 4 The diagram shows the results of processing IPIX data using the three methods of NSPPVD, NCWT, and RP in this invention.

[0041] Figure 5 This is the result of image fusion in this invention.

[0042] Figure 6 The images show the detection results of different detection methods of this invention.

[0043] Figure 7 This is a detection map of the Yantai sea area in an embodiment of the present invention. Detailed Implementation

[0044] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0045] Please see Figures 1-7 The present invention provides the following technical solutions:

[0046] This invention proposes a weak target detection method for the sea surface based on multimodal time-frequency image fusion. It converts a one-dimensional signal into three grayscale images with complementary characteristics using three signal processing methods: Normalized Wavelet Transform (NCWT), Normalized Smooth Pseudo-Wigner-Vell Transform (NSPWVD), and Recursive Graph (RP). The resulting two-dimensional grayscale images are then mapped to the R, G, and B channels using weighted coefficients for feature fusion, generating a fused color image. Subsequently, an improved multi-modal EfficientNetV2s target detection framework (IMMEV2s) is constructed, and weak signal detection with a controllable false alarm rate is achieved by optimizing the decision threshold of the Softmax layer.

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] like Figure 1 The method for detecting weak targets on the sea surface based on multimodal time-frequency image fusion is shown. This method includes three parts: data augmentation, image fusion, and model detection.

[0049] (1) Data Augmentation

[0050] To meet the false alarm rate requirements, a sufficient number of samples is necessary. Therefore, this invention uses a sliding window processing method with a step size of 64 and a window length of 1024 to process the sea clutter signal. For each dataset, more than 20,000 time signals of length 1024 are obtained. Each signal segment is processed using CWT, SPWVD, and RP to obtain three grayscale images. Normalization is performed on the CWT and SPWVD methods: pure clutter cells are selected as reference cells, and their mean and variance of CWT and SPWVD are calculated. This is then used to normalize all cells under test. Normalization enhances the power intensity of the target wave and suppresses the power intensity of clutter, thus improving image discriminability. The results are then grayscale mapped to obtain a TF grayscale image.

[0051] When generating TF maps, CWT can display the high-frequency transients of sea clutter, while SPWVD can accurately analyze the frequency modulation characteristics of sea clutter. Furthermore, RP is used for nonlinear time series analysis to reveal the dynamic invariance of the system (such as attractor structure). This method of joint analysis of multimodal time-frequency features achieves complementary physical dimensions and synergistic anti-interference capabilities, enriching image features and significantly improving the detection and analysis capabilities of weak signals in complex environments.

[0052] (2) Image Fusion

[0053] Three grayscale images generated through signal processing are weighted and mapped to the R, G, and B channels for feature fusion. The resulting color image is then scaled to a uniform size of 224*224 to match the tensor dimension requirements of the model input. This method preserves the time-frequency, energy, and dynamic characteristics of the signal, preventing information loss. Furthermore, multiple modal inputs avoid signal contamination of a single modality, improving anti-interference capabilities.

[0054] Compared to common image stitching methods, the method of this invention has a smaller input dimension, lower computational cost, clear physical meaning, intuitive color visualization, and is naturally compatible with CNN architecture, requiring no adjustment to the input layer structure, making it suitable for practical engineering applications.

[0055] (3) Model Detection

[0056] To better utilize the acquired multimodal images, this paper uses EfficientNetV2s as the backbone classification network and introduces an attention mechanism. It dynamically enhances the feature responses of important channels through convolutional layers and SE layers. The improved multimodal EfficientNetV2s detection model has the following structure diagram: Figure 2 As shown.

[0057] In radar target detection against sea clutter backgrounds, sample imbalance is a key challenge. Since the number of sea clutter samples (H0 hypothesis) is typically significantly greater than the number of true target samples (H1 hypothesis), the traditional cross-entropy loss function leads to a severe bias towards the majority class during model training, reducing the ability to detect weak targets and ultimately significantly increasing the probability of missed detections. To address this sample imbalance problem, Focal Loss is introduced to replace cross-entropy loss. By adjusting the focusing parameter γ and class weight α, the loss contribution from sea clutter is reduced, and the model's focus on weak targets is enhanced. The formula is as follows:

[0058] FL(pt)=-α(1-p t ) γ log(p t )

[0059] In the formula, p t This represents the model's predicted probability for the true class.

[0060] Furthermore, in the context of small target detection on the sea surface, the false alarm rate and detection probability are the two most important metrics. The false alarm rate is the proportion of all actual non-target samples that are incorrectly identified, while the detection probability is the proportion of target samples that are correctly identified. To avoid the model results being affected by imbalanced sample sizes, the false alarm rate is generally controlled at 10%. -3 .

[0061] To address the constant false alarm rate (CFAR) problem, this invention proposes a detection strategy based on softmax probability output and employs the Monte Carlo method to determine the decision boundary for a controllable CFAR. In a binary classification scenario, the softmax output reflects the probability of belonging to two categories, as shown in the following equation:

[0062]

[0063] The classifier will make judgments based on a preset threshold θ, which defaults to 0.5. The classification decision rules are as follows:

[0064]

[0065] In the formula, P(class1) is the probability that the model predicts a sample belongs to class 1. To control the false alarm rate, this threshold needs to be manually controlled; when the false alarm rate is below 10... -3 When the false alarm rate is higher than 10, lower the threshold to improve the model's detection probability; when the false alarm rate is higher than 10, lower the threshold to improve the model's detection probability. -3 At times, the threshold can be increased to reduce the false alarm rate. For example... Figure 3 As shown, by adjusting the decision threshold to the position indicated by the red line, the number of samples classified as "1" can be controlled, thereby making false alarms in target detection controllable.

[0066] This method uses the Dartmouth 1993 dataset from the IPIX radar. This data was collected from a cliff top near Dartmouth, Nova Scotia, on the east coast of Canada using the IPIX radar. The radar transmission frequency was 9.3 GHz, the pulse repetition frequency was 2000 Hz, the sampling distance interval was 15 m, and the target was a small sphere wrapped in aluminum wire with a diameter of 1 m. Data set information is shown in Table 1.

[0067] Table 1 IPIX Dataset Information

[0068]

[0069] In the table, SWH represents the significant wave height, and WS represents the wind speed. The table shows that the first eight datasets are in sea states 2-3, while the latter two are in sea states 3-4, with the tenth dataset having the highest significant wave height. Broken wave crests and scattering whitecaps in the waves can obscure target echoes, and small targets are sometimes easily blocked by waves. When the observation time is too short, it will be difficult to distinguish between sea crests and small targets.

[0070] The three signal processing methods of this invention are used to process a 1024-bit segment of the dataset to obtain an image as shown below. Figure 4 As shown, the result after image fusion is as follows: Figure 5 As shown.

[0071] Figure 4 As can be seen, for the NSPWVD method, the obtained target wave image has a high-power strip with zero frequency, and the power is more concentrated at the midline, with a higher overall power. In contrast, the power distribution of pure clutter is more uniform and lower. For the NCWT method, the brightness in the image reflects the changed power intensity. It can be seen that the TF image of the target wave has some high-power regions, while the TF image of pure clutter has relatively fewer high-power regions and a lower peak power. For the RP method, pure clutter has randomness and short-term correlation, and its RP image is represented by vertical and horizontal line segments. In contrast, the target echo has definite dynamic characteristics, and its RP image is represented by diagonal, spiral, or rotating structures. Figure 5 It is the result of weighted fusion of three methods, which can clearly distinguish the target signal from the pure clutter signal.

[0072] Under HH polarization, the detection probabilities of the proposed method are compared with those of three-feature detection methods, time-frequency three-feature detection methods, fractal detection methods, six-feature detection methods, and FAST four-feature detection methods. Figure 6As shown. The training set has 17,890 samples, and the test set has 4,473 samples. To ensure the false alarm rate is below 10⁻³, the number of false detections is controlled to be within 5. The number of false alarms is controlled at 4. If it is below 4, the threshold is lowered to increase the false alarm rate and improve the detection probability; if it is above 4, the threshold is raised to decrease the detection probability and control the false alarm rate.

[0073] The results show that the accuracy of the method of this invention is 93.94%, which is about 21% higher than that of existing detection methods. For the last six sets of data, due to the good sea conditions and high signal-to-clutter ratio, general time-frequency detection algorithms can achieve good results. However, for the Data26, Data30, and Data31 datasets, the accuracy of the detection algorithm is reduced due to the low signal-to-clutter ratio or high wind speed. For the Data17 dataset, although its signal-to-clutter ratio is very high, its effective wave height reaches 2.2m, the sea conditions are extremely poor, and the signal is easily blocked by waves, resulting in a relative decrease in the accuracy of the algorithm. The method of this invention combines three transformation methods, SPWVD, CWT, and RP, to enhance the features of the data and improve the classifier for prediction. Experimental results show that the proposed IMMEV2s detector outperforms the other four detectors, especially under low sea state conditions.

[0074] To evaluate the model's tolerance to disturbances such as noise and changes in input data quality, experiments were conducted using the Yantai dataset from China to test the model's robustness in complex environments. This dataset is from the "Radar Maritime Detection Data Sharing Project" of the Naval Aviation University. The latest data from the dataset, 20221112150043_stare_HH, was used. The radar model is Tian'ao SPPR50P, with polarization HH. The transmitted combined pulses consist of a single pulse signal T1, an LFM pulse signal T2, and an LFM pulse signal T3, with an overall repetition frequency of 2000Hz. The target is a steel light buoy at 2.97 nautical miles, with an effective wave height of 1.8m and a sea state of 4. Using the T2 single pulse signal, the target was detected as buoy 2. There are 1000 echo range gates, with the target range gate located at 666-674. The signal length is 131000, and the total observation time is 65.5s.

[0075] A segment of pure clutter data with a length of 1024 and an observation duration of 0.512s was selected, along with a segment of the target data. The three proposed methods were used to convert the sequence into an image, and the results are as follows. Figure 7 As shown.

[0076] Ten pure clutter range gates (specific values: 1, 100, 200, 250, 300, 350, 400, 850, 900, 950) and one buoy location range gate (specific value: 671) from the dataset were selected. The 11 resulting signals underwent the same sliding window processing as described above, yielding 22,363 sea clutter sequences of length 1024 and observation duration 0.512s. Using the method of this invention, these sequences were transformed from one-dimensional to two-dimensional, and the same number of color images were obtained through the RGB channels. The training and test sets were divided in an 8:2 ratio. After model training, the prediction set showed a Test Loss of 0.0024, a detection rate of 0.9720, and an accuracy of 0.9973. The resulting confusion matrix... Among them, 4043 and 418 are the number of actual sea clutter identified as sea clutter and the number of actual targets identified as targets, respectively; 12 and 0 are the number of actual targets identified as sea clutter and the number of actual sea clutter identified as targets, respectively.

[0077] At this point, the model's false alarm rate is 0. The detection probability can be increased by appropriately lowering the threshold of the Softmax layer, while ensuring the false alarm rate does not exceed 10⁻³. After lowering the threshold, a new confusion matrix is ​​obtained. The detection probability was increased to 0.9906. The results show that the method invented this year can effectively detect weak target signals on the sea surface of Yantai under the condition of constant false alarm rate.

[0078] In summary, this invention solves the problem of difficulty in detecting weak signals against a sea clutter background, and has good detection results for different sea conditions.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion, characterized in that, Includes the following steps: S1: Obtain the signal to be detected, which includes sea clutter signal and target echo signal. Divide the signal to be detected into training data and test data. Perform sliding window processing on the obtained data to obtain several signals to be detected. S2: For each segment of the signal to be measured, normalized smooth pseudo-Wigner transform, normalized continuous wavelet transform and recursive graph are applied to obtain three two-dimensional grayscale images with complementary properties. S3: The obtained two-dimensional grayscale image is mapped to the R, G, and B channels through weighted coefficients for feature fusion to generate a fused color image; S4: Construct an improved classification framework for EfficientNetV2s, input the obtained images into this classification framework, and detect weak signals; The improved EfficientNetV2s classification framework uses EfficientNetV2s as the backbone classification network, introduces an attention mechanism, dynamically enhances the feature responses of important channels through convolutional layers and SE layers, introduces Focal Loss to replace cross-entropy loss, and adjusts the focusing parameters. and category weight The formula for reducing the loss contribution of sea clutter and enhancing the model's attention to weak targets is as follows: In the formula, This represents the model's predicted probability for the true class.

2. The method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion according to claim 1, characterized in that, In step S1, the ratio of training data to test data is 7:3, the width of the sliding window is 1024, the delay is 64, and the length of the signal to be tested is 1024.

3. The method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion according to claim 1, characterized in that, In S2, the smoothed pseudo-Wigner transform SPWVD, for a given complex signal Its SPWVD formula is as follows: In the formula, It is a frequency smoothing window. It is a time-smoothing window that uses sea clutter cells as reference cells to calculate the mean value of its SPWVD. With variance For each unit under test, this mean and variance are used for normalization to obtain NSPWVD, as shown in the following formula: 。 4. The method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion according to claim 2, characterized in that, In S2, the mathematical definition of the continuous wavelet transform (CWT) is as follows: In the formula, For scale parameters, For translation parameters, It is the complex conjugate of the mother wavelet function. Morlet wavelet is selected as the mother wavelet, and clutter cells of sea clutter are used as reference cells to calculate the mean CWT. With variance For each unit under test, this mean and variance are used for normalization to enhance the time Doppler characteristics of sea clutter target echoes, resulting in the NCWT, as shown in the following formula: NCWT has higher frequency resolution for low-frequency components and higher time resolution for high-frequency components, and can effectively capture the local time-frequency characteristics of sea clutter signals.

5. The method for detecting weak targets on the sea surface based on multimodal time-frequency map fusion according to claim 1, characterized in that, In S2, the recursive graph process is as follows: Step 1: Select the appropriate embedding dimension With time delay The phase space reconstruction of the one-dimensional time series yields the following spatial vectors: Step 2: Calculate the Euclidean distance between each pair of trajectories after phase space reconstruction: Step 3: Construct the recursive matrix, the specific equation is as follows: In the formula, It is the distance threshold, which determines the criterion for similarity. It is a step function; Step 4: Generate a visual RP. By applying grayscale mapping, different distance values ​​in the distance matrix are mapped to different grayscale values, thereby obtaining a grayscale image that encapsulates the complete features of the time series and forming a texture map that intuitively displays the dynamic characteristics of the time series.

Images