A SAR raw echo target screening and coarse positioning method and system based on phase-aware hyper computation

Abstract

The application provides a SAR original echo target screening and coarse positioning method and system based on phase-aware hyper-computing, and relates to the technical field of real-time target detection of synthetic aperture radar. The method directly faces the complex SAR original echo, uses amplitude-phase decomposition, phase index hyper-vector dictionary, amplitude weighted binding accumulation, prototype hyper-vector Hamming matching and sliding window regional level response screening, realizes target existence screening and target coarse position estimation without complete imaging. The application can retain amplitude intensity information and phase structure information, reduce the calculation and storage cost before the imaging link, and improve the target / background separability under the condition of strong clutter and low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of real-time target detection technology of synthetic aperture radar, and in particular to a method and system for screening and coarsely locating SAR raw echo targets based on phase-sensing superdimensional calculation. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing technology with all-weather, day-and-night observation capabilities, widely used in scenarios such as patrol inspection, disaster emergency response, and ground feature monitoring. With the development of spaceborne and airborne SAR systems, platforms need to continuously acquire large-scale coherent raw echo data and complete target detection tasks within the constraints of limited downlink bandwidth, on-chip storage resources, and real-time processing capabilities. However, in actual observation scenarios, only a small number of echo areas are related to potential targets; most sampled data corresponds to background clutter or non-suspicious areas. Performing complete imaging and subsequent detection on the entire scene's raw echoes would incur significant computational, storage, and communication overhead.

[0003] Existing SAR target detection methods mainly include image domain detection, partial imaging or pre-imaging domain detection, and raw echo domain detection. Image domain methods typically rely on a complete imaging process including range compression, range migration correction, and azimuth compression, followed by target recognition using constant false alarm rate (CFAR) detection or convolutional neural networks. While offering high detection accuracy, these methods suffer from long computational chains and high hardware resource consumption. Partial imaging or pre-imaging domain methods can reduce the imaging burden to some extent, but still rely on range compression, focus-related preprocessing, or statistical clutter modeling. Their performance is susceptible to complex clutter, low signal-to-clutter ratio, and scene inhomogeneity. Deep learning-based methods possess strong feature extraction capabilities, but typically involve numerous floating-point convolution operations and high storage overhead, making them unsuitable for real-time screening on resource-constrained platforms such as spaceborne and airborne systems.

[0004] Hyperdimensional computing (HDC) represents data using high-dimensional bipolar hypervectors and performs similarity reasoning using operations such as binding, accumulation, symbolization, and Hamming distance comparison. It has advantages such as simple computation, storage friendliness, strong robustness, and ease of hardware implementation.

[0005] Chinese invention patent CN122063576A discloses a SAR target screening method based on binary superdimensional calculation. This method directly targets the raw SAR echo data and achieves target presence screening through I / Q dual-channel stitching, binarization quantization, adaptive segmentation, random binary projection, and Hamming distance comparison. It can quickly distinguish between background samples and suspected target samples without performing complete imaging, thereby reducing the front-end screening computational overhead.

[0006] However, the aforementioned prior applications still have room for further improvement: First, the method mainly performs hyperdimensional encoding based on the difference in amplitude after real, imaginary, or binarization, failing to fully utilize the local structural relationships and coherent features contained in the phase of the original SAR echo, which can easily lead to insufficient target-background separability in strong clutter and low signal-to-noise ratio scenarios; Second, the method uses symbolic quantization and random binary projection to generate sample hypervectors, which, although simple to implement in hardware, weakens the fine-grained information carried by continuous amplitude and phase changes to some extent, especially unfavorable for preserving the local coherent features of weak targets; Third, the method focuses on target presence screening, mainly outputting the judgment result of whether a target exists, and has not yet provided a rough position of the target in the original echo domain within the same lightweight hyperdimensional computation framework, thus making it difficult to directly provide candidate regions for subsequent local imaging, fine identification, or downlink data cropping.

[0007] Therefore, how to further establish an effective mapping relationship between phase and supervector on the basis of inheriting the low complexity advantage of the original echo domain superdimensional screening, jointly utilize amplitude intensity and phase consistency to enhance the separability of targets and clutter, and achieve target presence screening and coarse localization within the same lightweight framework, remains a key problem to be solved in the field of SAR front-end intelligent processing. Summary of the Invention

[0008] Purpose of the invention: To address the problems of existing SAR target detection relying on complete or partial imaging links, high computational overhead of front-end screening, insufficient utilization of phase structure, and difficulty in unifying screening and coarse localization within the same lightweight framework, this invention proposes a method and system for SAR raw echo target screening and coarse localization based on phase-aware hyperdimensional computation. This method directly targets complex SAR raw echoes and utilizes amplitude-phase decomposition, phase index hypervector dictionary, amplitude weighted binding accumulation, prototype hypervector Hamming matching, and sliding window region-level response screening to achieve target existence screening and coarse target position estimation without requiring complete imaging.

[0009] To achieve the above technical objectives, this invention proposes a method for screening and coarsely locating SAR raw echo targets based on phase-sensing hyperdimensional computation, comprising the following steps:

[0010] S1. Obtain the raw SAR echo complex matrix ,in For azimuth sampling index, For range-direction sampling index, the original SAR echo complex matrix Perform amplitude-phase decomposition to obtain the amplitude matrix A and the phase matrix. The amplitude matrix A is then normalized according to a preset amplitude range to obtain a normalized amplitude matrix. ; S2. Construct a phase index supervector dictionary, which includes multiple bipolar phase supervectors, each of which corresponds to a phase interval. S3, According to the phase matrix The phase value of each sampling point is queried from the phase index hypervector dictionary to obtain the corresponding phase hypervector, and then the normalized amplitude matrix is ​​used. The phase supervector is weighted by the amplitude values ​​of the corresponding sampling points. S4. Accumulate all weighted phase hypervectors and perform bipolarization processing to generate the query hypervector; S5. Compare the query hypervector with the background class hypervector and the target class hypervector obtained from offline learning, respectively, and output the target screening result based on the magnitude of the Hamming distance or the difference in the Hamming distance. S6. When the target screening result indicates that a target exists, a sliding window is extracted on the complex original echo matrix according to a preset window size and step size. Steps S1 to S4 are performed on each window to obtain the window query hypervector. A response map is generated based on the Hamming distance difference between the window query hypervector and the background class hypervector and the target class hypervector. S7. Perform regional filtering on the response map to determine the target coarse positioning area and output the target coarse positioning position.

[0011] As a preferred embodiment, in step S1, the amplitude matrix and phase matrix respectively satisfy... , , where ∠ represents calculation The phase magnitude of each complex element, ranging from The normalized magnitude matrix The element in the i-th row and j-th column ,in and These are the preset lower limit and the preset upper limit of amplitude, respectively. This represents the amplitude limiting function, which normalizes the amplitude matrix. Forced restrictions Interval.

[0012] As a preferred embodiment, in step S2, the phase index supervector dictionary is constructed in the following ways: Randomly generate initial phase hypervector ,in The dimension of the hypervector; According to phase interval Will The phase range is divided into K intervals, among which floor( () indicates rounding down; By randomly flipping the set of non-overlapping dimensions of the previous phase hypervector, the following generation is performed sequentially. , … This ensures that the phase supervectors corresponding to adjacent phase intervals have a similarity exceeding a preset value, and that the Hamming distance of the corresponding phase supervectors gradually increases as the phase difference increases.

[0013] As a preferred embodiment, in step S3, regarding the phase value : when When, map it to ,in ; when At that time, it is first shifted to the positive phase range and then its sign is inverted, mapped to ,in Therefore, only explicit storage is required. The range corresponds to the phase supervector.

[0014] As a preferred embodiment, in step S4, the query supervector ,in Let i be the phase hypervector corresponding to the sampling point in the i-th row and j-th column. The normalized amplitude corresponding to the sampling point in the i-th row and j-th column is... This indicates that bipolar symbolization is performed dimension-by-dimensionally on the accumulated result, resulting in elements with values ​​of... Bipolar query hypervector.

[0015] As a preferred embodiment, in step S5, the background class hypervector and the target class hypervector are obtained through offline learning, specifically including: Steps S2 to S5 are performed on the background training samples and the target training samples respectively to obtain the training sample hypervectors; The training sample hypervectors of the same category are accumulated and bipolarized in dimension to generate the initial category hypervectors; The initial class hypervector is retrained using training samples to correct misclassifications. When a training sample is misclassified, its hypervector is added to its true class hypervector and subtracted from the incorrectly predicted class hypervector to increase the Hamming distance margin between the background class and the target class.

[0016] As a preferred embodiment, step S5, which involves outputting the target screening result based on the Hamming distance magnitude or Hamming distance difference, specifically includes: Let h be the query hypervector and h be the background class hypervector. The Hamming distance between them is Query hypervector With target class hypervector The Hamming distance between them is ;when If the target is present in the complex original echo matrix, it is determined that the target exists; otherwise, it is determined to be a background scene.

[0017] As a preferred embodiment, in step S6, the window position in the response graph... The response value at that location is ,in Window position The corresponding window query hypervector, This represents the Hamming distance; the larger the response value, the greater the probability of the presence of a target in that window region.

[0018] As a preferred embodiment, in step S7, the region-level filtering includes: performing smoothing processing on the response map R to obtain a smoothed response map. According to the preset response threshold Extracting high-response regions Perform connectivity analysis on the high-response region to obtain multiple candidate connected regions; calculate the region score for each candidate connected region, which is related to the average response value and area of ​​the region; select the connected region with the largest region score as the target coarse localization region, and use the center of the connected region as the target coarse localization position.

[0019] Furthermore, this invention proposes a SAR raw echo target screening and coarse localization system, which is used to execute the aforementioned SAR raw echo target screening and coarse localization method based on phase-sensing hyperdimensional calculation. The system includes: The amplitude and phase preprocessing module is used to obtain the raw SAR echo complex matrix. ,in For azimuth sampling index, For range-direction sampling index, the original SAR echo complex matrix Perform amplitude-phase decomposition to obtain the amplitude matrix A and the phase matrix. The amplitude matrix A is then normalized according to a preset amplitude range to obtain a normalized amplitude matrix. ; The phase index supervector dictionary module is used to construct a phase index supervector dictionary, which includes multiple bipolar phase supervectors, each corresponding to a phase interval; based on the phase matrix... The phase value of each sampling point is queried from the phase index hypervector dictionary to obtain the corresponding phase hypervector, and then the normalized amplitude matrix is ​​used. The phase supervector is weighted by the amplitude values ​​of the corresponding sampling points. The hyperdimensional encoding module is used to accumulate all weighted phase hypervectors and perform bipolarization processing to generate the query hypervector; The prototype learning module is used to generate background class hypervectors and target class hypervectors; The target screening module is used to compare the query hypervector with the background class hypervector and the target class hypervector respectively using Hamming distance, and output the target screening result based on the magnitude of the Hamming distance or the difference in Hamming distance. The sliding window coarse localization module is used to perform sliding window truncation on the complex original echo matrix according to the target screening results and a preset window size and step size. For each window, a window query hypervector is calculated, and a response map is generated based on the Hamming distance difference between the window query hypervector and the background class hypervector and the target class hypervector. The response map is then filtered at the regional level to determine the target coarse localization region and output the target coarse localization position.

[0020] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention directly processes SAR complex raw echoes, and target screening and coarse positioning can be completed at the front end without performing a complete imaging process for the entire scene, which can significantly reduce the data scale entering the subsequent imaging or fine detection modules.

[0021] Second, this invention utilizes both amplitude and phase information. Amplitude is used to characterize echo intensity, and phase is used to index the structural hypervector, which can transform the local coherence of the target echo into similarity enhancement in a hyperdimensional space, thereby improving the separability of target clutter.

[0022] Third, the phase index dictionary of this invention maps sampling points with similar phases to similar hypervectors, and sampling points with large phase differences to approximately orthogonal hypervectors. Clutter and noise phases are relatively dispersed and tend to cancel each other out during the hyperdimensional binding accumulation process, while the target phase has strong local consistency and can be effectively enhanced.

[0023] Fourth, this invention employs operations such as bipolar supervectors, table lookup, weighted accumulation, symbolization, and Hamming distance comparison. The core process is hardware-friendly and suitable for lightweight implementation on space, air, or at the edge.

[0024] Fifth, after screening, the present invention further provides a sliding window coarse positioning function, which obtains the approximate location of the target through regional response selection, and can provide a priori area for subsequent fine imaging, target recognition or communication downlink cropping.

[0025] Sixth, this invention supports updating parameters such as amplitude normalization range, phase interval, hypervector dimension, window size, step size, and response threshold through training samples, which can adapt to different radar systems, different resolutions, and different sea states / ground features. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0027] Figure 2 This is a schematic diagram showing the distribution of different data representation methods for raw SAR echoes.

[0028] Figure 3 This is a schematic diagram illustrating the construction of the phase index supervector and the phase mapping relationship of the present invention.

[0029] Figure 4 A comparison chart showing the detection accuracy of different hyperdimensional encoding methods under different hypervector dimensions.

[0030] Figure 5 This is a schematic diagram of the coarse localization response results based on a sliding window under different signal-to-noise ratio conditions. Detailed Implementation

[0031] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0032] This embodiment is designed for front-end target screening tasks on spaceborne or airborne SAR platforms, with the input being complex raw echo data directly acquired by SAR sensors. In this embodiment, the raw SAR echo matrix to be processed is denoted as... ,in This is the azimuth sampling index, with a value range of [value range missing]. , This is the distance sampling index, with a value range of [value range missing]. Each element in the matrix is ​​a complex echo sample value, containing the scattering amplitude and phase information at that sampling location, and the corresponding observation scene size is... The target size is set to Radar system center frequency signal bandwidth Pulse duration Distance sampling rate Platform height Slant distance from scene center Platform speed Antenna length Azimuth sampling rate or pulse repetition frequency Operating wavelength The parameters above are for illustrative purposes only; in actual applications, they can be configured according to the radar system parameters.

[0033] like Figure 1As shown, this embodiment includes an offline learning stage and an online inference stage. The offline learning stage is used to generate a phase index hypervector dictionary and prototype hypervectors for the background class and the target class; the online inference stage is further divided into two parts: target screening and sliding window coarse localization, which are used to perform fast screening on the input raw echo and output the coarse localization region when needed.

[0034] (a) Offline learning stage The purpose of the offline learning phase is to establish a stable hyperdimensional representation foundation for subsequent online inference. This mainly includes amplitude-phase preprocessing, phase-indexed hypervector dictionary construction, training sample hyperdimensional encoding, category hypervector generation, and category hypervector retraining optimization. In this embodiment, the training set contains 729 pure background samples and 729 samples containing targets, and the signal-to-noise ratio range covered by the training samples is [range missing]. ,Include , , , , Five typical signal-to-noise ratio conditions. The test set contains 225 pure background samples and 225 samples containing the target, used to validate screening and coarse localization performance.

[0035] First, amplitude-phase decomposition is performed on the SAR complex raw echoes acquired during the training phase. For any training sample... Calculate the magnitude matrix respectively and phase matrix ,in This represents the amplitude response at each sampling point in the original echo. This indicates the phase of each sampling point. The amplitude mainly reflects the response intensity of the target or background scattered echo, while the phase includes the local structural relationship and coherence characteristics of the echo. Figure 2 This paper presents the distribution characteristics of raw SAR echoes containing a central target under four representations: real part, imaginary part, normalized amplitude, and phase. It can be seen that amplitude more directly reflects the target response area, while phase, like the real and imaginary parts, contains certain positional and structural variation trends. Therefore, this invention employs a joint representation of amplitude and phase in the hyperdimensional encoding process, rather than relying solely on amplitude information.

[0036] To avoid excessive differences in the dynamic range of amplitudes across different scenarios affecting the subsequent hyperdimensional accumulation process, the amplitude matrix needs to be normalized. In this embodiment, the lower limit of amplitude normalization is set to... The upper limit of amplitude normalization is set to For the th element in the normalized magnitude matrix... Line number Column elements ,in This represents the amplitude limiting function, used to limit the normalized amplitude to a certain value. Within this range. After this processing, target-related sampling points with larger amplitudes have higher weights in subsequent hyperdimensional accumulation, while the impact of abnormal amplitudes or extreme noise on the final hypervector is effectively limited.

[0037] After obtaining the amplitude and phase representations of the training samples, a phase hypervector dictionary is further constructed. In this embodiment, the hypervector dimension is set to... The phase quantization interval is set to Therefore, the positive phase range The number of phase hypervectors that need to be constructed is .

[0038] To construct the phase hypervector dictionary, an initial phase hypervector is first randomly generated: Then, An generates them sequentially according to the phase interval order. , … This makes the Hamming distance between the phase supervectors corresponding to adjacent phase intervals only 0. ,exist , Under typical values, the similarity between adjacent phase hypervectors exceeds 90%, and the Hamming distance between corresponding phase hypervectors gradually increases as the phase difference increases. For example... Figure 3 As shown, adjacent phase hypervectors are obtained by randomly flipping non-overlapping dimensions. In this embodiment, each time an adjacent phase hypervector is generated, it is randomly flipped... The non-reused dimension maps sampling points with similar phases to similar hypervectors, while sampling points with large phase differences are mapped to hypervectors with low similarity. In this way, the continuous phase change relationship in the complex domain is transformed into a similarity change relationship in the hyperdimensional space, thus enabling the phase structure to participate in subsequent target-background differentiation.

[0039] For any sampling point in the training samples, its phase value is denoted as... This invention employs different mapping methods based on the phase sign. When When, directly look up the corresponding phase hypervector based on the phase interval index; when At that time, first add the phase value The vector is shifted to the positive phase range, and then the sign of the found phase supervector is inverted. The specific mapping relationship is as follows: ,in , This mapping method only requires explicit storage. The 180 phase hypervectors within the range are used to obtain the corresponding representation for negative phases through phase shifting and sign inversion. This reduces the storage overhead of the phase dictionary and makes it easier to deploy on hardware platforms with limited on-chip storage resources.

[0040] After constructing the phase-indexed hypervector dictionary, the system can encode training samples as sample hypervectors in a hyperdimensional space. For the Line number The phase hypervector of the column sampling points is denoted as... The normalized amplitude is denoted as This invention uses normalized magnitude as the weight for... The phase hypervectors corresponding to each sampling point are weighted and accumulated, and the accumulation result is subjected to dual polarization processing to obtain the sample hypervector. ,in This indicates that bipolar symbolization is performed dimension-by-dimensionally on the accumulated result, resulting in elements with values ​​of... The bipolar query hypervector is used. In this encoding process, the normalized amplitude determines the contribution of each sampling point to the final hypervector, while the phase hypervector determines the structural position of the sampling point in the hyperdimensional space. For target-related regions, since their amplitudes are usually strong and the local phases have a certain consistency, multiple target-related sampling points will be mapped to similar phase hypervectors and will be enhanced during high-dimensional accumulation. For background clutter or random noise, since their phase distribution is relatively dispersed, different sampling points are mapped to multiple different phase hypervectors, which are easily canceled out during accumulation in high-dimensional space.

[0041] Let the background training sample set be The target training sample set is The above encoding process was performed on both the background training samples and the target training samples, resulting in 729 background training hypervectors. and 729 target training hypervectors ,in This describes the hyperdimensional encoding process of the amplitude-weighted phase hypervector described in this invention. Subsequently, training hypervectors of the same class are accumulated dimensionally and subjected to bipolarization processing to obtain the initial background class hypervector. and the initial target class hypervector Through this process, background samples and target samples form corresponding category prototypes in hyperdimensional space. Subsequent online inference can then complete target screening by comparing the Hamming distance between the input sample hypervector and the category prototype.

[0042] To further improve the separability between the background class and the target class, this embodiment retrains the initial class hypervectors 20 times. For any training sample hypervector... Its real label is The system determines the predicted label based on the Hamming distance between the training sample and the current class hypervector. ,in This indicates the calculation of Hamming distance. Indicates the first During the second training iteration, the class Category hypervectors, If the predicted label With real labels If they are consistent, it means that the current category hypervector can correctly represent the training sample, and the category hypervector remains unchanged; if the predicted label With real labels If there is an inconsistency, the misclassified sample is used to correct the class hypervector. Specifically, the sample hypervector is added to its true class hypervector, and the sample hypervector is subtracted from the incorrectly predicted class hypervector. Then, the bipolarization process is re-executed on the updated result. , .

[0043] After 20 rounds of retraining, the final background class supervector is obtained. and target class hypervector These two categories of hypervectors are stored in associative memory, on-chip memory, or local memory, serving as reference prototypes for the target screening and sliding window coarse localization stages.

[0044] (II) Target Screening Phase The target screening stage is used to determine the presence of targets in the input raw SAR echoes. This stage operates directly on the complex raw echoes and does not require the generation of a complete SAR image first. Therefore, it can be used as a pre-screening module in spaceborne, airborne, or edge-end SAR processing links.

[0045] During online inference, the system receives the raw complex echo of the SAR to be detected. And normalize the range according to the amplitude determined in the offline learning phase. , Phase quantization interval Number of phase indices and hypervector dimension The same encoding process is performed. First, amplitude-phase decomposition is performed on the original input echo to obtain the amplitude matrix. and phase matrix Then, the amplitude matrix is ​​normalized to obtain the normalized amplitude matrix. Subsequently, the phase value of each sampling point is quantized in intervals, and the corresponding phase supervector is obtained by querying the phase index supervector dictionary based on the phase sign.

[0046] For the Line number The column sampling points have the following phase hypervector: The normalized amplitude is The system uses normalized amplitude as weight to perform weighted summation of the phase hypervectors of all sampling points, and then performs dual polarization processing to generate the query hypervector corresponding to the input sample. .

[0047] Get the query hypervector Then, the system calculates its hypervectors relative to the background class. and target class hypervector Hamming distance between and ,when If the input sample is closer to the target class prototype in the hyperdimensional space, the system determines that there is a target in the original SAR echo. This indicates that the input sample is closer to the background class prototype, and the system determines that the original SAR echo is a targetless or background scene.

[0048] Figure 4 Accuracy comparisons of various HDC coding methods under different hypervector dimensions are presented. In this experiment, the hypervector dimension increased from 1000 to 10000, with intervals of 1000. It can be seen that as the hypervector dimension increases, the overall accuracy of various HDC coding methods improves. However, the phase-aware coding method used in this invention outperforms traditional coding methods such as Level-ID and random projection in most dimension settings. This indicates that by establishing a structured mapping relationship between phase and hypervector, and combining it with amplitude-weighted accumulation, the separability between the target and the background in the raw SAR echo can be effectively enhanced.

[0049] (III) Sliding Window Coarse Positioning Stage When the target screening stage determines that a target exists in the input SAR raw echo, the system further enters the sliding window coarse localization stage. The purpose of this stage is not to generate a high-resolution SAR image, but to quickly determine the approximate area where the target may appear in the raw echo domain, thereby providing prior information for subsequent local imaging, fine detection, or data downlink cropping.

[0050] First, in the original echo matrix A sliding window is set at the top. In this embodiment, the height of the sliding window is set to... The width of the sliding window is set to The azimuth sliding step size is set to The distance sliding step size is set to For each window position The system extracts local echo blocks. ,in This indicates a partial window clipping operation, and the result of the clipping. For each local window The system repeatedly performs the same amplitude-phase decomposition, amplitude normalization, phase lookup table, amplitude weighted accumulation, and bipolarization processing as in the target screening phase to obtain the window query hypervector. Since each window only covers a local region of the original echo, the query hypervector of this window can characterize the similarity between the corresponding local region and the target class or background class.

[0051] Subsequently, the system calculates the window query hypervectors respectively. Hypervectors of the background class Target class hypervector The Hamming distance between them is calculated, and the Hamming distance difference response is constructed. This response value reflects whether the current window is closer to the background class or the target class. When the window mainly contains background clutter, the window hypervector is usually closer to the background class hypervector. Smaller window hypervectors result in lower response values; when the window contains target-related echoes, the window hypervector is closer to the target class hypervector. Smaller response values ​​result in higher response values. Therefore, the response graph obtained after traversing all windows... It can be used to indicate the location area where a target may appear.

[0052] To avoid positioning instability caused by isolated noise peaks under conditions of strong clutter or low signal-to-noise ratio, this invention does not directly select a single maximum response point, but instead employs a region-level screening strategy. In this embodiment, the response map is first... Perform mean smoothing to obtain a smoothed response map. Then, based on the preset threshold Extracting high-response regions

[0053] Next, connectivity analysis was performed on the high-response regions to obtain several candidate connected regions. ,in, This indicates the connected component extraction operation. This represents the number of candidate connected regions. For each candidate region... Calculate its regional score ,in, This represents the average response value within the candidate region. This represents the area of ​​the candidate region. The candidate region with the highest region score is ultimately selected as the coarse localization region for the target. The center of the area is used as the coarse location result of the target.

[0054] Figure 5 Coarse localization response results based on a sliding window are presented under different signal-to-noise ratios (SNRs). In this experiment, the SNRs were as follows: , , , , The target size is The target is located in the center of the scene. It can be seen that under low signal-to-noise ratio (SNR) conditions, several clutter-related high-response regions may still exist in the response map, but a distinguishable regional response can still be formed near the target; as the SNR increases, the response in the central target region becomes more concentrated, and the coarse target localization result becomes more stable. Experimental results show that... Under the condition that the center positioning error is Under other conditions, the center positioning errors are respectively , , and This result demonstrates that the present invention can not only perform target presence screening at the raw echo level, but also provide coarse target location information without performing complete imaging.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A SAR raw echo target screening and coarse positioning method based on phase-aware hyper-computation, characterized in that, Includes the following steps: S1, obtaining a SAR original echo complex matrix wherein is an azimuth sampling index, is a range sampling index, performing amplitude-phase decomposition on the SAR original echo complex matrix to obtain an amplitude matrix A and a phase matrix , and normalizing the amplitude matrix A according to a preset amplitude range to obtain a normalized amplitude matrix ; S2. Construct a phase index supervector dictionary, which includes multiple bipolar phase supervectors, each of which corresponds to a phase interval. The phase index hypervector dictionary is constructed in the following ways: Randomly generating an initial phase super-vector wherein is the super-vector dimension; According to phase interval Will The phase range is divided into K intervals, among which floor( () indicates rounding down; By randomly flipping the set of non-overlapping dimensions of the previous phase hypervector, the following generation is performed sequentially. , … This makes the phase supervectors corresponding to adjacent phase intervals have a similarity exceeding a preset value, and the Hamming distance of the corresponding phase supervectors gradually increases as the phase difference increases; S3, According to the phase matrix The phase value of each sampling point is queried from the phase index hypervector dictionary to obtain the corresponding phase hypervector, and then the normalized amplitude matrix is ​​used. The phase supervector is weighted by the amplitude values ​​of the corresponding sampling points. S4. Accumulate all weighted phase hypervectors and perform bipolarization processing to generate the query hypervector; S5. Compare the query hypervector with the background class hypervector and the target class hypervector obtained from offline learning, respectively, and output the target screening result based on the magnitude of the Hamming distance or the difference in the Hamming distance. S6. When the target screening result indicates that a target exists, a sliding window is extracted on the SAR raw echo complex matrix according to a preset window size and step size. Steps S1 to S4 are performed for each window to obtain the window query hypervector, and a response map is generated based on the Hamming distance difference between the window query hypervector and the background class hypervector and the target class hypervector. S7. Perform regional filtering on the response map to determine the target coarse positioning area and output the target coarse positioning position.

2. The SAR raw echo target screening and coarse localization method based on phase-aware hyperdimensional calculation according to claim 1, characterized in that, In step S1, the amplitude matrix and phase matrix respectively satisfy... , , where ∠ represents calculation The phase magnitude of each complex element, ranging from The normalized magnitude matrix The element in the i-th row and j-th column ,in and These are the preset lower limit and the preset upper limit of amplitude, respectively. This represents the amplitude limiting function, which normalizes the amplitude matrix. Forced restrictions Interval.

3. The SAR raw echo target screening and coarse localization method based on phase-aware hyperdimensional calculation according to claim 1, characterized in that, In step S3, regarding the phase value : when When, map it to ,in ; when At that time, it is first shifted to the positive phase range and then its sign is inverted, mapped to ,in Therefore, only explicit storage is required. The range corresponds to the phase supervector.

4. The SAR raw echo target screening and coarse localization method based on phase-aware high-dimensional calculation according to claim 1, characterized in that, In step S4, the query supervector ,in Let i be the phase hypervector corresponding to the sampling point in the i-th row and j-th column. The normalized amplitude corresponding to the sampling point in the i-th row and j-th column is... This indicates that bipolar symbolization is performed dimension-wise on the accumulated result, resulting in elements with values ​​of... Bipolar query hypervector.

5. The SAR raw echo target screening and coarse localization method based on phase-aware high-dimensional calculation according to claim 1, characterized in that, In step S5, the background class hypervector and the target class hypervector are obtained through offline learning, specifically including: Steps S2 to S5 are performed on the background training samples and the target training samples respectively to obtain the training sample hypervectors; The training sample hypervectors of the same category are accumulated and bipolarized in dimension to generate the initial category hypervectors; The initial class hypervector is retrained using training samples to correct misclassifications. When a training sample is misclassified, its hypervector is added to its true class hypervector and subtracted from the incorrectly predicted class hypervector to increase the Hamming distance margin between the background class and the target class.

6. The SAR raw echo target screening and coarse localization method based on phase-aware hyperdimensional calculation according to claim 4, characterized in that, In step S5, the step of outputting the target screening result based on the Hamming distance or the Hamming distance difference specifically includes: Let h be the query hypervector and h be the background class hypervector. The Hamming distance between them is Query hypervector With target class hypervector The Hamming distance between them is ;when If the target is present in the original SAR complex echo matrix, it is determined that a target exists; otherwise, it is determined to be a background scene.

7. The SAR raw echo target screening and coarse localization method based on phase-aware high-dimensional calculation according to claim 1, characterized in that, In step S6, the window position in the response graph The response value at that location is ,in Window position The corresponding window query hypervector, This represents the Hamming distance; the larger the response value, the greater the probability of the presence of a target within that window region.

8. The SAR raw echo target screening and coarse localization method based on phase-aware high-dimensional calculation according to claim 1, characterized in that, In step S7, the region-level filtering includes: performing smoothing processing on the response map R to obtain a smoothed response map. According to the preset response threshold Extracting high-response regions Perform connectivity analysis on the high-response region to obtain multiple candidate connected regions; calculate the region score for each candidate connected region, which is related to the average response value and area of ​​the region; select the connected region with the largest region score as the target coarse localization region, and use the center of the connected region as the target coarse localization position.

9. A SAR raw echo target screening and coarse localization system, used to execute the SAR raw echo target screening and coarse localization method based on phase-sensing superdimensional calculation as described in any one of claims 1 to 8, characterized in that, The system includes: The amplitude and phase preprocessing module is used to obtain the raw SAR echo complex matrix. ,in For azimuth sampling index, For range-direction sampling index, the original SAR echo complex matrix Perform amplitude-phase decomposition to obtain the amplitude matrix A and the phase matrix. The amplitude matrix A is then normalized according to a preset amplitude range to obtain a normalized amplitude matrix. ; The phase index supervector dictionary module is used to construct a phase index supervector dictionary, which includes multiple bipolar phase supervectors, each corresponding to a phase interval; based on the phase matrix... The phase value of each sampling point is queried from the phase index hypervector dictionary to obtain the corresponding phase hypervector, and then the normalized amplitude matrix is ​​used. The phase supervector is weighted by the amplitude values ​​of the corresponding sampling points. The hyperdimensional encoding module is used to accumulate all weighted phase hypervectors and perform bipolarization processing to generate the query hypervector; The prototype learning module is used to generate background class hypervectors and target class hypervectors; The target screening module is used to compare the query hypervector with the background class hypervector and the target class hypervector respectively using Hamming distance, and output the target screening result based on the magnitude of the Hamming distance or the difference in Hamming distance. The sliding window coarse localization module is used to perform sliding window truncation on the SAR raw echo complex matrix according to the target screening results and according to the preset window size and step size. For each window, the module calculates the window query hypervector and generates a response map based on the Hamming distance difference between the window query hypervector and the background class hypervector and the target class hypervector. The module performs regional-level filtering on the response map to determine the target coarse localization area and outputs the target coarse localization position.

Images