Radar target template signal establishment method based on deep model self-adaptive segmentation

Abstract

The application discloses a radar target template signal establishment method based on deep model adaptive segmentation, mainly relates to the technical field of template signal, and aims to solve the problem that the existing simple threshold slicing is used for binary segmentation of a time-frequency graph, which is easy to cause target contour fracture or false detection by clutter, and the threshold parameter is very sensitive to signal-to-clutter ratio and environmental changes. The method comprises the following steps: reading a one-dimensional radar echo sequence, mapping to a two-dimensional time-frequency domain to obtain a time-frequency image; obtaining a multi-scale feature and a prompt matrix; obtaining a binary mask corresponding to a mask feature through the multi-scale feature and the prompt matrix; determining that a binary mask corresponding to a highest confidence score is a final mask; screening time-frequency transform data corresponding to the radar echo sequence by using the final mask to obtain output data; and restoring the output data to a time-domain signal by using an inverse short-time Fourier transform, and taking the time-domain signal as a target template signal.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of template signal, in particular to a radar target template signal establishment method based on deep model adaptive segmentation. BACKGROUND

[0002] When performing radar target detection and parameter estimation in sea clutter or ground clutter environment, constructing a pure target reference template signal is an important task. The template needs to completely retain the Doppler variation and amplitude fluctuation of the target, while eliminating clutter, noise and interference as much as possible, so as to improve the separability of the target and the background.

[0003] With the development of digital signal processing technology, time-frequency analysis tools such as Short Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) have been widely used in radar preprocessing. However, relying only on simple threshold slicing to perform binary segmentation on time-frequency graph is easy to cause target contour to break or be misdetected by clutter, and the threshold parameter is very sensitive to signal to clutter ratio (SCR) and environmental changes, which is difficult to ensure the stability and reliability of the template. SUMMARY

[0004] The present application provides a radar target template signal establishment method based on deep model adaptive segmentation to solve the problem that the existing scheme uses simple threshold slicing to perform binary segmentation on time-frequency graph, which is easy to cause target contour to break or be misdetected by clutter, and the threshold parameter is very sensitive to signal to clutter ratio (SCR) and environmental changes, which is difficult to ensure the stability and reliability of the template.

[0005] In a first aspect, the present application provides a radar target template signal establishment method based on deep model adaptive segmentation, the method comprising:

[0006] Reading a one-dimensional radar echo sequence, mapping to a two-dimensional time-frequency domain to obtain a complex matrix; after time-frequency transformation and pseudo-color mapping of the complex matrix, obtaining a time-frequency image;

[0007] Loading the pre-training weight of SAM2, calling the image encoder to perform multi-scale processing on the input time-frequency image to obtain multi-scale features;

[0008] According to the prompt point coding injected by the time-frequency image, a prompt matrix is obtained;

[0009] Through the multi-scale features and the prompt matrix, a plurality of mask features are obtained, and then a binary mask corresponding to the mask features is obtained;

[0010] Calculate the confidence scores corresponding to the mask features, and determine the binary mask corresponding to the highest confidence score as the final mask;

[0011] Using the final mask, the time-frequency transform data corresponding to the radar echo sequence is filtered to obtain the filtered time spectrum. The time spectrum is restored to the original frequency arrangement order through frequency inverse rearrangement to obtain the output data. The output data is restored to the time domain signal using inverse short-time Fourier transform, and the time domain signal is used as the target template signal.

[0012] In one implementation of this application, a one-dimensional radar echo sequence is read, mapped to a two-dimensional time-frequency domain, and a complex matrix is ​​obtained; the complex matrix is ​​then subjected to time-frequency transformation and pseudo-color mapping to obtain a time-frequency image, specifically including:

[0013] Reading one-dimensional radar echo sequences Wherein, the sampling rate of the radar echo sequence is equal to the pulse repetition frequency, denoted as . ;

[0014] Using preset formulas:

[0015] By mapping the radar echo sequence to the two-dimensional time-frequency domain, a complex matrix is ​​obtained. ;

[0016] in, For length is Hamming window function, and The first The first time window center and the first Each frequency point, The total length of the signal is denoted by m; the value of m ranges from [1, M], where M represents the total number of preset frequency points.

[0017] Through the formula

[0018] , obtain along The frequency axis performs a centered rearrangement of the time-frequency transformation data;

[0019] in, This indicates a frequency rearrangement operation;

[0020] Through the formula Take the complex spectrum Logarithmic compression is performed on the amplitude to obtain compressed data;

[0021] Where ε is a preset small constant to prevent underflow caused by a logarithm of zero;

[0022] Will Linear normalization to the (0,1) interval;

[0023] By Jet Pseudo Color Function :

[0024] ,

[0025] mapping the radar echo sequence to a three-channel color space to obtain a time-frequency image;

[0026] wherein, ⌊⋅⌋ is the floor function.

[0027] In an implementation manner of the present application, the pre-training weight of SAM2 is loaded, and the image encoder is called to perform multi-scale processing on the input time-frequency image to obtain multi-scale features, specifically including:

[0028] The height and width of the time-frequency image are set as and The image encoder will output a new feature map after each downsampling, and finally obtain multi-scale features ;

[0029] wherein, ,

[0030] is the scale index, is the feature tensor of the th layer, is the total number of layers, is the number of channels of the th layer.

[0031] In an implementation manner of the present application, the prompt point coding according to the time-frequency image is obtained to obtain a prompt matrix, specifically including:

[0032] Marking P groups of positive and negative prompt points on the time-frequency image to form a set ;

[0033] wherein, , is the pixel coordinate of the th prompt point, is a preset label, 1 representing a foreground point and 0 representing a background point;

[0034] The query vector

[0035] is obtained by calculation through the formula: ; wherein, is a preset mapping length;

[0036] All are vertically spliced to obtain a prompt matrix .

[0037] In one implementation of this application, several mask features are obtained through multi-scale features and a cue matrix, and then a binary mask corresponding to the mask features is obtained, specifically including:

[0038] The mask decoder extracts image features With hint matrix Generate through cross-attention fusion Group preset low-resolution mask features ,index ;

[0039] The original time-frequency image size is sampled from each set of mask features and then processed by the Sigmoid function. Activation yields a pixel-level probability mask. ;

[0040] right Generate a binary mask with a threshold of 0.5 .

[0041] In one implementation of this application, the confidence score corresponding to the mask feature is calculated, and the binary mask corresponding to the highest confidence score is determined as the final mask, specifically including:

[0042] Through the formula:

[0043] Calculate the confidence score for each mask feature. ;

[0044] in, , and It refers to the linear layer weights and biases. Indicates global average pooling;

[0045] from The binary mask corresponding to the mask feature with the highest confidence score among the mask features is selected as the final mask.

[0046] In one implementation of this application, the time-frequency transform data corresponding to the radar echo sequence is filtered using a final mask to obtain the filtered time spectrum. The time spectrum is then restored to its original frequency arrangement order via frequency inverse rearrangement to obtain output data. The output data is then restored to a time-domain signal using an inverse short-time Fourier transform, and this time-domain signal is used as the target template signal. Specifically, this includes:

[0047] Through the formula:

[0048] Obtain the filtered time spectrum;

[0049] in, represents the time-frequency transformed data after centering rearrangement along the frequency axis of the complex matrix;

[0050] The output data is obtained by the formula:

[0051]

[0052] wherein, represents the inverse centering processing of the frequency dimension;

[0053] The output data is restored to a time-domain signal by using the inverse short-time Fourier transform, and finally the target template signal is obtained .

[0054] From the above technical solutions, the present application has the following advantages:

[0055] The present application realizes multi-scale feature processing of time-frequency images by loading the image encoder of the SAM2 pre-training weight. Compared with the contour breaking problem caused by the traditional threshold slicing method, the present application can adaptively capture multi-level information from local micro-Doppler features to global motion trajectories. The multi-scale feature fusion mechanism effectively overcomes the limitations of single-scale analysis on complex moving targets, especially for dynamic targets with harmonic characteristics such as helicopter rotors and vehicle vibrations, which can maintain their physical continuity. This feature extraction method based on deep learning fundamentally avoids the sensitivity problem caused by manually setting thresholds.

[0056] The prompt point encoding mechanism introduced in the present application converts the prior knowledge of the target region of the operator into a mathematical prompt matrix, and generates a binary mask in cooperation with the multi-scale features. This design has double advantages: first, it significantly reduces the false detection rate of clutter regions by guiding attention focusing through human-computer interaction; second, the confidence score mechanism automatically selects the optimal mask to ensure that the filtered time-frequency spectrum has the highest signal-to-clutter ratio. Compared with the sensitivity of traditional schemes to environmental changes, this data-driven and expert knowledge combination method makes the template construction still maintain stable output in a strong clutter background.

[0057] The frequency inverse rearrangement technology used in the final output stage solves the problem of physical information distortion caused by frequency axis rearrangement in the time-frequency transformation process. By strictly restoring the original Doppler frequency arrangement order, in cooperation with the inverse short-time Fourier transform, it ensures that the generated time-domain template signal not only retains the intrinsic motion characteristics of the target (such as micro-Doppler modulation period), but also eliminates incoherent noise. This closed-loop processing process makes the template signal truly reflect the kinematic characteristics of the target, providing a physically interpretable input source for subsequent recognition and classification. BRIEF DESCRIPTION OF DRAWINGS

[0058] ​In order to more clearly illustrate the technical solutions of the present application, the drawings required to be used in the description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0059] Figure 1 is a radar target template signal establishment method flowchart based on a deep model adaptive segmentation provided by an embodiment of the present application.

[0060] Figure 2 is a raw signal time-frequency image provided by an embodiment of the present application.

[0061] Figure 3 is an interactive point labeling image provided by an embodiment of the present application.

[0062] Figure 4 is an automatic segmentation mask image provided by an embodiment of the present application.

[0063] Figure 5 is a target energy band reservation image provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort fall within the protection scope of the present application.

[0065] Those skilled in the art should understand that the embodiments described below are only preferred embodiments of the present disclosure, and do not mean that the present disclosure can only be implemented by the preferred embodiments. The preferred embodiments are only used to explain the technical principles of the present disclosure, and are not used to limit the protection scope of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by those skilled in the art without creative effort still fall within the protection scope of the present disclosure.

[0066] It should also be noted that the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or further include elements inherent in such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0067] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0068] The embodiment provides a method for establishing radar target template signals based on depth model adaptive segmentation, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0069] Step 110: Read the one-dimensional radar echo sequence, map it to the two-dimensional time-frequency domain to obtain a complex matrix; after time-frequency transformation and pseudo-color mapping of the complex matrix, obtain the time-frequency image.

[0070] In some embodiments, this step may specifically be as follows:

[0071] Reading one-dimensional radar echo sequences Wherein, the sampling rate of the radar echo sequence is equal to the pulse repetition frequency, denoted as . ;

[0072] Using preset formulas:

[0073] By mapping the radar echo sequence to the two-dimensional time-frequency domain, a complex matrix is ​​obtained. ;

[0074] in, For length is Hamming window function, and The first The first time window center and the first Each frequency point, The total length of the signal is denoted by m; the value of m ranges from [1, M], where M represents the total number of preset frequency points.

[0075] Through the formula

[0076] , obtain along The frequency axis performs a centered rearrangement of the time-frequency transformation data;

[0077] in, This indicates a frequency rearrangement operation;

[0078] Through the formula Take the complex spectrum Logarithmic compression is performed on the amplitude to obtain compressed data;

[0079] Where ε is a preset small constant to prevent underflow caused by a logarithm of zero;

[0080] Will Linear normalization to the (0,1) interval;

[0081] By Jet Pseudo Color Function :

[0082] ,

[0083] mapping the radar echo sequence to a three-channel color space to obtain a time-frequency image (as shown in Figure 2 );

[0084] wherein ⌊⋅⌋ is a floor function, (k, m) is an image pixel coordinate, corresponding to a signal discrete index .

[0085] Step 120, load the pre-training weight of SAM2, call the image encoder to perform multi-scale processing on the input time-frequency image to obtain multi-scale features.

[0086] Loading the pre-training weight of SAM2, calling the image encoder to perform multi-scale processing on the input time-frequency image to obtain multi-scale features, specifically including:

[0087] Setting the height and width of the time-frequency image as and , the image encoder will output a new feature map after each downsampling, and finally obtain multi-scale features ;

[0088] wherein, ,

[0089] is a scale index, is the feature tensor of the layer, is the total number of layers, is the number of channels of the layer.

[0090] It should be noted that all layers collectively provide multi-layer information from fine texture to large structure, laying a foundation for subsequent mask generation.

[0091] Step 130, obtaining a prompt matrix according to the prompt point coding injected by the time-frequency image.

[0092] According to the prompt point coding injected by the time-frequency image (as shown in Figure 3 ), a prompt matrix is obtained, specifically including:

[0093] Marking P groups of positive and negative prompt points on the time-frequency image to form a set ;

[0094] wherein, , is the pixel coordinates of the prompt points, is a preset label, 1 represents a foreground point, and 0 represents a background point;

[0095] is obtained by the formula:

[0096] ; wherein, is a preset mapping length (herein: is normalized to , so that the coordinates are aligned with the network input scale, and then mapped to a query vector with a length of by two layers of perception machines );

[0097] vertically spliced to obtain a prompt matrix .

[0098] Step 140, a plurality of mask features are obtained by the multi-scale feature and the prompt matrix, and a binary mask corresponding to the mask feature is further obtained.

[0099] In some embodiments, a plurality of mask features are obtained by the multi-scale feature and the prompt matrix, and a binary mask corresponding to the mask feature is further obtained, specifically including:

[0100] The mask decoder fuses the image feature and the prompt matrix through cross-attention to generate a set of preset low-resolution mask features , indexed as ;

[0101] Each group of mask features is upsampled to the original time-frequency image size and activated by a Sigmoid function to obtain a pixel-level probability mask ;

[0102] wherein, ;

[0103] The binary mask is generated with 0.5 as the threshold . .

[0104] Step 150, a confidence score corresponding to the mask feature is calculated, and a binary mask corresponding to the highest confidence score is determined as the final mask.

[0105] The confidence score corresponding to the mask feature is calculated, and a binary mask corresponding to the highest confidence score is determined as the final mask, specifically including:

[0106] ​​Through the formula:

[0107] Calculate the confidence score for each mask feature. It should be noted that, The larger the value, the higher the probability that the mask matches the real foreground.

[0108] in, , and It refers to the linear layer weights and biases. Indicates global average pooling;

[0109] from The binary mask corresponding to the mask feature with the highest confidence score among the mask features is selected as the final mask. The automatic segmentation mask image corresponding to the final mask (final veneer) is shown below. Figure 4 As shown.

[0110] Step 160: Using the final mask, filter the time-frequency transform data corresponding to the radar echo sequence to obtain the filtered time spectrum. The time spectrum is restored to the original frequency arrangement order through frequency inverse rearrangement to obtain the output data. The output data is restored to the time domain signal using the inverse short-time Fourier transform, and the time domain signal is used as the target template signal.

[0111] This step specifically includes:

[0112] Through the formula:

[0113] Obtain the filtered time spectrum;

[0114] in, This represents the time-frequency transform data after centering and rearranging along the frequency axis of the complex matrix.

[0115] Through the formula:

[0116] Obtain the output data;

[0117] in, This indicates inverse centralization of the frequency dimension;

[0118] The output data is restored to a time-domain signal using the inverse short-time Fourier transform, and the target template signal is finally obtained. The target energy band retention diagram corresponding to the target template signal is shown below. Figure 5 As shown.

[0119] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for establishing a radar target template signal based on deep model adaptive segmentation, characterized in that, The method comprises: reading a one-dimensional radar echo sequence, mapping to a two-dimensional time-frequency domain to obtain a complex matrix; after time-frequency transformation and pseudo-color mapping of the complex matrix, a time-frequency image is obtained; loading the pre-training weight of SAM2, calling the image encoder to perform multi-scale processing on the input time-frequency image to obtain multi-scale features; obtaining a prompt matrix according to the prompt point coding injected by the time-frequency image; obtaining a plurality of mask features through the multi-scale features and the prompt matrix, and then obtaining a binary mask corresponding to the mask features; calculating the confidence score corresponding to the mask features, and determining the binary mask corresponding to the highest confidence score as the final mask; using the final mask to filter the time-frequency transformation data corresponding to the radar echo sequence to obtain filtered time-frequency spectrum, and the time-frequency spectrum is restored to the original frequency arrangement order through frequency inverse rearrangement to obtain output data; using inverse short-time Fourier transform to restore the output data to a time-domain signal, and taking the time-domain signal as a target template signal.

2. The radar target template signal establishment method based on deep model adaptive segmentation according to claim 1, characterized in that, reading a one-dimensional radar echo sequence, mapping to a two-dimensional time-frequency domain to obtain a complex matrix; obtaining a time-frequency image after time-frequency transformation and pseudo-color mapping of the complex matrix, specifically comprising: Reading a one-dimensional radar echo sequence wherein the sampling rate of the radar echo sequence is equal to the pulse repetition frequency, denoted as ; using a preset formula: mapping the radar echo sequence to a two-dimensional time-frequency domain to obtain a complex matrix ; wherein, is a Hamming window function with a length of , and are the center of the mth time window and the kth frequency point, respectively, , , is the total length of the signal; m is in the range of [1, M], and M represents the total number of preset frequency points; through the formula, , obtaining along frequency axis performing the centering rearranged time-frequency transform data; wherein denotes a frequency rearrangement operation; The formula is , and the amplitude of the complex spectrum is logarithmically compressed to obtain compressed data. wherein ε is a preset small constant to prevent overflow caused by zero logarithm; will be described in the following with reference to the accompanying drawings. In the drawings: linearly normalized to the interval (0, 1); By Jet Pseudo Color function : ， mapping the radar echo sequence to a three-channel color space to obtain a time-frequency image; wherein ⌊⋅⌋ is the floor function.

3. The method of claim 1, wherein, loading the pre-training weight of SAM2, calling the image encoder to perform multi-scale processing on the input time-frequency image to obtain multi-scale features, specifically comprising: The height and width of the time-frequency image are set to and The image encoder outputs a new feature map after each downsampling, and finally obtains multi-scale features ; wherein , is a scale index, is a first layer feature tensor, is a total number of features, is a first is a number of channels of the layer.

4. The method of claim 1, wherein, obtaining a prompt matrix according to the prompt point coding injected by the time-frequency image, specifically comprising: Marking P sets of positive and negative hint points on the time-frequency image, forming a collection ; wherein, , is the pixel coordinate of the first prompt point, is the pixel coordinate of the first prompt point, is a preset label, 1 represents a foreground point, and 0 represents a background point. through the formula: , the query vector is obtained by calculation ; wherein, is a preset mapping length; all vertically spliced to obtain a prompt matrix .

5. The method of claim 1, wherein, obtaining a plurality of mask features through the multi-scale features and the prompt matrix, and then obtaining a binary mask corresponding to the mask features, specifically comprising: The mask decoder decodes the image features with the prompt matrix to generate a set of preset low-resolution mask features indexed ; Up-sampling to the original time-frequency image size on each group of mask features and passing through a Sigmoid function Activation to get pixel-level probability masks ; To Generate a binary mask with 0.5 as the threshold .

6. The method of claim 1, wherein, calculating the confidence score corresponding to the mask features, and determining the binary mask corresponding to the highest confidence score as the final mask, specifically comprising: through the formula: computing a confidence score for each mask feature ; wherein, , and are linear layer weights and bias, denotes global average pooling; From The binary mask corresponding to the mask feature with the highest confidence score is selected as the final mask from the mask features.

7. The method of claim 1, wherein, using the final mask to filter the time-frequency transformation data corresponding to the radar echo sequence to obtain filtered time-frequency spectrum, and the time-frequency spectrum is restored to the original frequency arrangement order through frequency inverse rearrangement to obtain output data; using inverse short-time Fourier transform to restore the output data to a time-domain signal, and taking the time-domain signal as a target template signal, specifically comprising: through the formula: obtaining the time-frequency spectrum after the screening; wherein, represents a time-frequency transformed data after a centering rearrangement along the frequency axis of the complex matrix; through the formula: obtaining output data; wherein denotes a frequency-wise inverse centering process; The output data is restored to a time domain signal by using inverse short-time Fourier transform, and finally the target template signal is obtained .

Images