A SAR target screening method based on binary hyper-computing

Abstract

The application provides a SAR target screening method based on binary hyper-computation, and relates to the technical field of real-time target detection. The method directly carries out real and imaginary part decomposition and sign function quantization on I / Q double-channel complex data of synthetic aperture radar (SAR) original echo, and obtains a binary fusion matrix; an abnormal profile is constructed based on a preset background statistical model, and the quantized data is adaptively divided into multiple segments according to the abnormal profile; each segment is subjected to hyper-dimensional mapping through an independent random binary projection matrix, segmented hyper-vectors are generated and spliced into complete sample hyper-vectors; and finally, target discrimination is realized by calculating the Hamming distance between the hyper-vectors and pre-stored background / target prototype hyper-vectors. The application can dynamically switch the quantization logic according to the target size, effectively solving the problem of poor adaptability of a single quantization strategy. Through an energy-aware adaptive segmentation mechanism, the weak target representation capability is effectively enhanced, and the screening accuracy in a strong clutter background is improved.

Description

Technical Field

[0001] This invention relates to the field of real-time target detection technology, and in particular to a SAR target screening method based on binary hyperdimensional calculation. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave remote sensing device, has the advantage of operating in all weather conditions and is widely used in target detection and ground feature monitoring. However, traditional SAR target detection relies on a complex imaging processing flow, which requires multiple steps such as range compression, azimuth compression, multi-look suppression, and coherent superposition. This involves multiple Fast Fourier Transform (FFT), interpolation, and phase compensation operations, resulting in a lengthy algorithm chain, high computational complexity, and high hardware implementation costs. At the same time, there is also a mismatch between imaging quality and downstream target detection tasks. Some detection scenarios do not require high-resolution focused images but still require the complete imaging process, resulting in a waste of computational resources. Moreover, target signals in the raw SAR echo data are often submerged by massive background noise, especially the echo signals of small targets such as small ships and aircraft, which are even weaker and difficult to distinguish from noise. In addition, the echo characteristics of targets of different sizes vary greatly, and a single processing strategy cannot take into account the discrimination performance of various targets. These problems together restrict the efficiency and applicability of traditional SAR target detection. Summary of the Invention

[0003] Purpose of the invention: To address the problems of traditional SAR target detection, which relies on complex imaging links, has high computational complexity, is easily overwhelmed by noise for small targets, and lacks adaptability to multiple target sizes, this invention proposes a SAR target screening method based on binary hyperdimensional computation (HDC). HDC replaces the traditional imaging process by directly performing real and imaginary partial solutions, adaptive quantization, and segmented hyperdimensional encoding on the raw SAR echo I / Q dual-channel complex data. Finally, by calculating the Hamming distance between the sample hypervector and the pre-stored background / target prototype hypervector, this invention achieves SAR target screening with low computational complexity, high noise robustness, and compatibility with multiple target sizes.

[0004] To achieve the above technical objectives, this invention proposes a SAR target screening method based on binary hyperdimensional calculation, comprising the following steps: S1. Obtain the raw SAR echo I / Q dual-channel complex data, and concatenate the real and imaginary matrices to generate a fused data matrix. S2. Based on the pre-stored original echo data of the noisy background, calculate the mean absolute value of the difference between each column in the fused data matrix and the corresponding background column to obtain the column intensity vector; S3. Accumulate all columns of the fused data matrix using a dynamic segmentation algorithm to obtain the cumulative sum of column strengths, and divide the fused data matrix into K segments based on the cumulative sum of column strengths, so that the cumulative sum of column strengths of each fused data sub-matrix tends to be equal. S4. Apply an independent random binary projection submatrix to each segment of the fused data submatrix to perform hyperdimensional projection, and obtain the floating-point projection results of each segment. S5. Determine the target size type based on the cumulative sum of column strengths described in step S3: If the cumulative sum of column strengths is less than a preset threshold, then perform sign function quantization on the floating-point projection results of each segment before accumulating them; otherwise, first accumulate the floating-point projection results of each segment before performing sign function quantization, and finally generate the sample hypervector. S6. Calculate the Hamming distance between the sample hypervector and the pre-stored background reference hypervector. If the Hamming distance is less than or equal to the preset judgment threshold, it is determined that there is no target; otherwise, it is determined that there is a target.

[0005] As a preferred approach, in step S1, the real and imaginary parts of the original SAR echo I / Q dual-channel complex data are extracted respectively, forming two dimensions. The real-valued matrix is ​​concatenated along the column direction to form a single matrix, generating a matrix of dimension . The fused data matrix is ​​constructed using index rearrangement. This stitching operation is achieved without introducing any floating-point operations or additional activation functions, preserving the complete phase and amplitude information of the original echo signal and providing structured input for subsequent binarization processing.

[0006] As a preferred embodiment, in step S2, the pre-stored original echo data of the noise background is obtained by accumulating multiple frames of SAR echo samples in a targetless scene and taking the average value, which is denoted as the background reference matrix; for each column in the fused data matrix, the absolute difference between it and the corresponding column element of the background reference matrix is ​​calculated, and the average value of all rows in the column is calculated to obtain a column intensity vector of length 2N, which is used to characterize the degree of deviation of each column from the background, and serves as the basis for subsequent adaptive segmentation.

[0007] As a preferred embodiment, in step S3, the dynamic segmentation algorithm includes: Calculate the cumulative sum of column strengths and iterate through the range of candidate segment number K. For each candidate segment number K, divide the 2N columns into K segments according to the threshold proportional to the cumulative sum of column strengths, so that the cumulative sum of column strengths in each segment tends to be equal. Calculate the Hamming distance difference between the hypervectors generated by the target sample and the background sample under the current segmentation scheme, select K that maximizes the difference as the optimal number of segments, and record the start and end column indices and segment lengths of each segment to form a fixed segmentation configuration for use in the inference stage.

[0008] As a preferred embodiment, step S4 specifically includes: Each fused data submatrix is ​​independently multiplied by its corresponding random binary projection submatrix. This random binary projection submatrix is ​​pre-generated during the training phase, with its elements independently and identically distributed in {-1, +1}, and has a dimension of 1. ,in The hyperdimensional dimension assigned to the i-th segment; The column number of the i-th segment is the segment length; the projection result is a segment. A 3D floating-point vector is used only for intermediate representation and is not involved in final storage or transmission.

[0009] As a preferred embodiment, step S5 specifically includes: Compare the cumulative sum of column strengths of the fused data matrix with a preset target size threshold: If the cumulative sum of column strengths is less than the preset threshold, it is determined to be a small target (weak targets such as 1×1 to 5×5 and no target scenarios). The projection results of each segment are first converted into a binary matrix by applying a sign function to each element, then accumulated by column to obtain an integer vector, and finally the sign is taken again to generate the final sample hypervector. If the cumulative sum of column strengths is greater than or equal to a preset threshold, it is determined to be a large target (such as 10×10 and above). First, the column strengths of each segment are accumulated by floating point to obtain a single vector, and then a sign function is applied uniformly for quantization.

[0010] As a preferred embodiment, in step S6, the pre-stored background reference hypervector is generated by performing majority voting on the resulting sample hypervectors after executing steps S1 to S5 on target-free scene samples. During inference, the Hamming distance between the query sample hypervector and the background reference hypervector is calculated. If the distance is less than or equal to a preset judgment threshold, a target is determined to exist; otherwise, it is determined to be pure background. This judgment logic avoids introducing a target prototype and achieves highly robust screening by relying solely on the background model.

[0011] Furthermore, the present invention also proposes an electronic device, which includes a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the above-described SAR target screening method based on binary hyperdimensional calculation.

[0012] Furthermore, the present invention also proposes a computer-readable storage medium storing at least one executable instruction, which, when executed on an electronic device, causes the electronic device to perform the aforementioned SAR target screening method based on binary hyperdimensional calculation.

[0013] Compared with the prior art, the present invention has the following beneficial effects: First, the HDC-based SAR raw echo data encoding algorithm of the present invention, through the collaborative design of dual-channel column splicing and difference mean adaptive segmentation module, accurately adapts to the complex dual-channel characteristics and target-noise strong coupling characteristics of SAR echo, effectively avoids signal cancellation and target energy dilution, and improves target feature recognition.

[0014] Second, the encoding algorithm of the present invention adopts an adaptive quantization strategy, which dynamically switches the quantization logic according to the target size, taking into account both the enhancement of weak signals of small targets and the preservation of amplitude information of large targets, thus solving the problem of poor adaptability of a single quantization strategy.

[0015] Third, the encoding algorithm of this invention only involves low-complexity operations such as array concatenation, mean calculation, matrix multiplication, accumulation and symbolic operation, without the need for complex operations such as FFT and phase compensation, which greatly reduces the computational overhead and adapts to the real-time processing needs of edge devices.

[0016] Fourth, the encoding algorithm of this invention leverages the characteristics of random noise canceling each other out and effective information enhancement in hyperdimensional space to significantly improve noise resistance in low signal-to-noise ratio scenarios, and is especially suitable for small-sized target detection.

[0017] Fifth, the coding algorithm of this invention adopts a modular design for each module, supports parameter configuration and function expansion at the software level, and can flexibly adapt to the SAR echo data processing needs of different frequency bands and different scenarios, with strong practicality and scalability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the SAR target screening method based on binary hyperdimensional calculation according to the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the application of quantization to a one-dimensional simple echo in the embodiment.

[0020] Figure 3 The Hamming distance relationship between sample hypervectors containing targets of different sizes and background class hypervectors under different data preprocessing and stitching strategies provided in the embodiments.

[0021] Figure 4 Hamming distance heatmap between a pure background and sample hypervectors containing targets of different sizes, provided for an embodiment.

[0022] Figure 5 The Hamming distance relationship between sample hypervectors containing targets of different sizes and background class hypervectors under different segmentation strategies provided in the embodiments. Detailed Implementation

[0023] 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.

[0024] Hyperdimensional computation (HDC), as an emerging computing framework, maps input data into hypervectors in a hyperdimensional space and uses Hamming distance to measure similarity, offering advantages in hardware friendliness and low complexity. This application applies HDC to the SAR target screening stage, proposing an adaptive binary hyperdimensional computation method as a preliminary coarse-screening step for target detection. This method directly performs real and imaginary partial solutions, binarization quantization, and energy-sensing segmentation on the raw SAR echo data, rapidly generating sample hypervectors and determining target presence, filtering over 90% of pure background samples without executing the complete imaging process. Through this screening mechanism, only potential target areas are retained, significantly reducing the amount of raw data that needs to be transmitted to the ground station and substantially reducing communication bandwidth pressure. Simultaneously, the input data size of subsequent high-precision target detection modules can be significantly reduced, greatly alleviating the computational burden, providing an efficient and feasible solution for real-time SAR target processing on spaceborne / airborne platforms.

[0025] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0026] Example 1: This embodiment uses the detection of small targets by SAR at sea as the application scenario, specifically including small ships, low-altitude aircraft and other typical small targets at sea. The raw SAR echo data is collected in the X band and dual polarization mode. The I / Q dual channel data scale is set to 1000×1024, with rows corresponding to the azimuth direction and columns corresponding to the range direction. Each row of complex vectors is all the sampled data of SAR at a point. After data stitching, 1000×2048 fused data is obtained.

[0027] The publicly disclosed SAR target screening process includes three stages: data preprocessing, training, and inference. Each stage is as follows: Figure 1 As shown. The preprocessing stage presets the core training parameters as follows: hypervector dimension. The value is 10000, used to determine the number of training samples for the segment length and dimension allocation scheme. The dataset is 1000, covering various maritime target sizes: no target, 1×1, 2×2, 3×3, 4×4, 5×5, 10×10, 20×20, 30×30, and 50×50. Each size contains 100 samples. The raw echo data for no target is first averaged and used as a reference background during the training phase. (Number of segments) The candidate values ​​range from 4 to 32, and the optimal solution is obtained through iteration. The threshold for determining the target size is set to the sum of column strengths. The value is 30. Characteristic optimization is performed on signals from small maritime targets. When the total strength of the columns is less than... Determined to be a small target, larger than The target is classified as large. Hamming distance threshold. The value is 2000, which is normalized to 0.2. This means that a normalized Hamming distance greater than 0.2 indicates the presence of a target, while a value less than or equal to 0.2 indicates the absence of a target. The background hypervector is generated offline in the final stage of training, using the reference background as input and the optimal encoding algorithm obtained through traversal. This generated hypervector serves as the reference hypervector in the inference stage and supports dynamic updates to adapt to different sea state scenarios.

[0028] The following sections will explain the three stages of data preprocessing, training, and inference in detail.

[0029] (a) Data preprocessing stage First, raw SAR echo data is acquired and preprocessed. Raw echo data of a maritime target scene is collected using a SAR sensor, outputting I / Q dual-channel complex data. Both the I and Q channels are 1000×1024, with a 32-bit floating-point format, containing both real and imaginary parts, denoted as . and ,in The value range is 0-999. The value range is 0-1023. The row index corresponds to the azimuth direction, i.e., the SAR sampling point. The column index corresponds to the range direction. Subsequently, 100 sets of SAR echo data for targetless marine scenes were collected, with each set having the same data size as the original echo. The I / Q channels of all samples were accumulated bit-by-bit and the mean was calculated, ultimately yielding a representative raw echo data set of the noise background, which was stored in a local cache for subsequent column intensity calculations. It should be noted that marine noise background exhibits spatiotemporal fluctuations; the statistical period of the 100 targetless samples covers typical scenarios under different sea conditions, thus ensuring that the noise background data reflects the noise distribution characteristics in practical applications.

[0030] After acquiring the raw echo data, it directly enters the fusion processing stage of the data stitching module. The data stitching module reads the preprocessed I / Q dual-channel raw data and employs a dual-channel column stitching strategy to reassemble the data according to column index order. I channel column indices 0-1023 correspond to fused data column indices 0-1023, and Q channel column indices 0-1023 correspond to fused data column indices 1024-2047. During the stitching process, data reassembly is achieved through array index mapping, eliminating the need for complex arithmetic operations and ensuring rapid processing at the edge. After stitching, a 1000×2048 fused data matrix is ​​generated and stored as a set of column vectors, each column being a 1000-dimensional vector, which is directly transmitted to the column strength calculation unit of the difference mean adaptive segmentation module. This column splicing method has been verified in multiple rounds and can effectively avoid the signal cancellation problem of I / Q channels of small targets at sea caused by row splicing. It fully preserves the weak scattering characteristics of small targets, especially the signal integrity protection effect of small targets such as 1×1 and 5×5. Through the complementary fusion of dual-channel data, the distinction between targets and noise is naturally improved.

[0031] (II) Training Phase The training phase primarily relies on the adaptive segmentation module for mean difference. The column intensity calculation unit receives the set of column vectors from the fused data, combines it with pre-stored ocean noise background data from the preprocessing phase, and calculates the sum of the absolute values ​​of the differences column-by-column, using the following formula: ,in The value range is 0-2047. Then, the column strength value for each column is calculated using the following formula: Ultimately, 2048 column strength values ​​were obtained, forming a column strength vector. The data is then transmitted to the dynamic segmentation unit. During the column intensity calculation, the differences between the fused data and the noise background are directly compared to fully explore the energy concentration characteristics of small maritime targets in the column dimension, providing a precise basis for subsequent segmentation and dimension allocation.

[0032] Dynamic segmentation unit loads column strength vector, traversing Given four candidate values: 4, 8, 16, and 32, for each... Perform the following operations on the input column intensity vector. The 2048 columns of data were then divided into segments based on the cumulative sum of column strength. Segments are created such that the cumulative column strength of each segment tends to be equal, and the length of each segment is denoted as . It is important to emphasize that during the segmentation process, it is necessary to strictly ensure that the cumulative column strength of each segment tends to be equal. This equalization strategy can prevent the target energy from being overly concentrated in a certain segment or split into multiple segments, ensuring the rationality of subsequent dimension allocation. Especially for the characteristic of small target energy being concentrated in a few columns, this segmentation scheme can accurately align the core area of ​​the target signal.

[0033] The dimension allocation unit then receives the data from each segment and the column strength vector, and calculates the mean segment strength for each segment using the following formula: Then calculate the sum of the average segment strength and the proportion of each segment. The total hypervector dimension is allocated to each segment according to its proportion. Dimensions assigned to each segment Thus, various The dimension allocation scheme corresponding to the value.

[0034] Here, the number of segments is chosen to balance the differences and the complexity of the encoding operations. =8, at this time the segment lengths corresponding to each segment are: =584、 =76、 =29、 =18、 =20、 =31、 =79、 =1211, and =2048, this segmentation method can accurately align the concentrated areas of small-sized targets, avoiding the energy of small-sized targets being diluted by noise. The segmentation scheme is stored in the module configuration file, and the 8 segments of data are extracted as 1000× The matrix is ​​used to synchronously transmit the column index range and segment length information of each segment to the dimension allocation unit and the hyperdimensional projection module. The corresponding dimension allocation is as follows: =84、 =641、 =1674、 =2801、 =2437、 =1672、 =643、 =48, and =10000, the core logic of which is to allow the target feature enrichment segments to obtain more dimensional resources, such as segments with higher average strength. , They were assigned to dimensions 2801 and 2437 respectively, while those with lower mean intensity... , The initial allocation is only 84-dimensional and 48-dimensional, ensuring that resources are tilted towards areas with concentrated effective information. Through the differentiated allocation of dimensional resources, the characteristic differences between the target and noise are further amplified.

[0035] Finally, after segmentation and dimension allocation, the proposed number of segments K=8 and the length of each segment are determined. Dimension allocation results The segmented column index range is packaged and transmitted to the hyperdimensional projection module. At the same time, the average segment strength and dimension allocation rules are stored in the cache for reuse in the inference stage. During inference, the pre-stored rules are directly called without repeated calculation.

[0036] (III) Reasoning Stage The inference phase revolves around the collaborative operation of the adaptive quantization module and the hyperdimensional projection module. Through a complete process of target size discrimination, segmented hyperdimensional projection, variable precision cumulative quantization, and similarity matching, the segmented fused data is transformed into target class hypervectors, ultimately achieving the existence detection of small targets at sea.

[0037] The adaptive quantization module first reuses the column strength calculation unit from the training phase and calculates the sum of column strengths of the input samples. This sum directly reflects the overall energy intensity of the target signal in the input sample. (The sum of column intensities is then used.) With the preset intensity threshold Compare and classify the target sizes. If Less than If the input sample is determined to be a small target, such as 1×1 or 5×5, or even a sample without a target, it will be classified as a small target. These targets have weak signals and highly concentrated energy, therefore quantization is performed before inputting them into the hyperdimensional projection module. Greater than Large targets, such as 10×10, 20×20, and 30×30, are identified as having richer signal amplitude information and relatively dispersed energy distribution. Therefore, the original data is directly input into the hyperdimensional projection module. For small targets, whose signals are weak and concentrated, the binarization unit prioritizes quantization of the input data, i.e., sign-taking operations. This process accumulates and strengthens the weight of the weak signals of small targets, preventing them from being submerged by ocean noise. For large targets, whose signal amplitude information is even richer, quantization operations would lose detailed information. Therefore, the target amplitude information is preserved, and the target class vector is obtained through subsequent accumulation and sign-taking operations.

[0038] The next step involves inputting the data processed by the adaptive quantization module into the hyperdimensional projection module. This module has been pre-loaded with a complete set of configuration parameters optimized during the training phase, including the segmentation scheme and dimension allocation scheme corresponding to the optimal number of segments. Simultaneously, a pre-stored 2048×10000 hyperdimensional matrix is ​​invoked, where each element is ±1. The matrix is ​​then divided into 8 non-overlapping sub-matrices according to the segmentation scheme. The segmentation projection principle and matrix dimension matching relationship are as follows: Figure 1 As shown in (c), the size of each submatrix is ​​precisely matched with the length and allocation dimension of the corresponding segment; that is, the size of the submatrix is ​​strictly equal to... This ensures complete data dimensional compatibility and no information loss during the projection process. After the projection operation is initiated, the input 8 segments of fused data, each segment having a size of [missing information], are processed. The hyperdimensional projection module performs matrix multiplication operations segment by segment, and the operation formula is as follows: ,in It refers to the subarray extracted from the hyperdimensional matrix. In the specific calculation process, element-wise multiplication is first performed on each element of each data segment with the corresponding element of the submatrix. Then, the multiplication results of each row are summed to obtain the projection result of that segment. Its scale is This projection process leverages the inherent characteristics of hyperspace to transform the sparse scattering features of small maritime targets into a distributed representation within hyperspace. Simultaneously, it utilizes the noise cancellation property of hyperspace to further suppress marine noise interference, exhibiting a significant enhancement effect, particularly on small target signals in low signal-to-noise ratio scenarios. After all eight data segments have been projected, the projection results of all segments are stitched together column-wise to form a complete projection matrix with a total size of 1000×10000.

[0039] The complete projection matrix output by the hyperdimensional projection module is then input into the adaptive quantization module, which performs corresponding accumulation and binarization operations based on the previous target size discrimination results. For small targets identified earlier, since pre-quantization processing has already been completed, the accumulation unit directly performs 8-bit integer accumulation on the projection matrix column-by-column to obtain a 1×10000 integer vector. Then, the binarization unit performs a sign function operation to convert the integer vector into a binary vector, which is the final target class hypervector. For large targets identified earlier, the accumulation unit first performs 32-bit single-precision floating-point accumulation on the projection matrix column-by-column to fully preserve the target amplitude details. Then, the binarization unit performs a sign operation on the floating-point accumulation result to generate a 1×10000 binary target class hypervector, completing the final hyperdimensional encoding. This encoding result serves as the core input for subsequent similarity matching, realizing the complete transformation from raw SAR echo data to a high-resolution hypervector.

[0040] The generated sample hypervectors are compared with the pre-stored background reference hypervectors from the training phase to calculate their similarity. Hamming distance is used as the core metric; a smaller Hamming distance indicates higher similarity, while a larger value indicates a more significant difference. The calculated Hamming distance is then used to... Compare with the preset judgment threshold of 2000: If This indicates that the original SAR echo sample contains small maritime targets. The sample was determined to contain only ocean noise background and no target was found. Through this similarity matching process, the final determination from the hyperdimensional encoding result to the target detection conclusion was completed, realizing end-to-end target detection of SAR raw echo without the need for traditional imaging procedures.

[0041] The underlying logic of the methods disclosed in the above embodiments can be used to build a SAR target screening system. This system integrates a data stitching module, an adaptive quantization module, a difference mean adaptive segmentation module, and a hyperdimensional projection module. These modules work together to extract and enhance features from the raw SAR echo data, providing a highly recognizable hyperdimensional vector representation for subsequent similarity matching.

[0042] The input of the data stitching module receives I / Q dual-channel data from the raw SAR echo, and its data size is [data size missing]. The output is connected to the input of the difference mean adaptive segmentation module; the data splicing module adopts a dual-channel column splicing strategy, splicing the N columns of data from channel I and the N columns of data from channel Q column by column, finally obtaining a data of size [missing information]. The fused data.

[0043] The difference mean adaptive segmentation module is the core module for adapting to the sparse distribution characteristics of raw SAR echoes. It integrates column intensity calculation unit, dynamic segmentation unit, and dimension allocation unit, and optimizes the number of segments offline using pre-trained samples. With the dimensionality allocation strategy, high-accuracy encoding configuration can be achieved with only a small number of training samples.

[0044] The column intensity calculation unit first acquires preset noise background data, which is obtained from statistical analysis of SAR echo samples from targetless scenes and includes the noise distribution characteristics of both I / Q channels. Then, it iterates through the sample data column by column, calculating the difference between each column of data and the corresponding column of noise background. ,in That is, the first The sum of the absolute values ​​of the differences in the column. For the sample data Line number The value of the column, This corresponds to the background noise value. Then, the mean of the absolute values ​​of the differences in each column is calculated, which gives the final column intensity value. .

[0045] In a further embodiment, the dynamic segmentation unit is randomly selected. A representative set of samples, covering different target sizes and scene noise levels; for the pre-trained sample set, traversal... Value, cumulative sum of column strength Column data splitting Segments, such that the cumulative column strength of each segment tends to be equal. For each candidate... The segmented results corresponding to the values, combined with subsequent hyperdimensional projection and quantization processes, are used to calculate the Hamming distance between the sample target class vector and the corresponding class vector in the hyperdimensional vector library, and to statistically analyze the differences. Hamming distance dissimilarity at a given value. Select the value with the largest Hamming distance dissimilarity. The value is used as the optimal number of segments, and the value is recorded at the same time. The segment length of each segment. This forms a fixed segmentation scheme, which is stored in the module's built-in configuration file and called directly during the inference phase.

[0046] In a further embodiment, the dimension allocation unit receives the segmentation strategy from the dynamic segmentation unit and the column strength value from the column strength calculation unit, and calculates the average column strength of all columns in each segment to obtain... Average strength of each segment , … Then, the proportion of the average segment strength to the total average difference of all segments was calculated. And allocate the hypervector dimension to each segment according to this ratio. , obtain the dimension allocation results For the initial dimension allocation results, the detection accuracy and Hamming distance difference of the pre-trained samples are calculated in conjunction with the complete encoding process. If the Hamming distance difference of a certain segment corresponding to the target is small, some dimensions are finely adjusted from the segments with smaller average differences and added to that segment until the detection performance is optimal. Finally, the optimized dimension allocation results are bound and stored with the segmentation scheme. During the inference phase, a dedicated hypervector dimension is allocated to each segment of data according to this scheme.

[0047] The input of the hyperdimensional projection module is connected to the output of the difference mean adaptive segmentation module, and its output is connected to the input of the adaptive quantization module. This module contains a built-in randomly generated hyperdimensional bipolar matrix with a size of [missing information]. To satisfy the rules of matrix multiplication, it can be directly applied to... The fused data undergoes full-dimensional projection. During the projection process, this module receives segmentation information output by the difference mean adaptive segmentation module, i.e., the length of each segment. This relates to dimensional allocation information, specifically the projected dimensions of each segment. and from Segmentally extract the corresponding submatrices from the hyperdimensional bipolar matrix. For the... Segment length is From the echo data, extract the corresponding column index range for that segment. The submatrix is ​​used to perform segmented projection by multiplying the submatrix with the corresponding fused data, so that each segment of data is mapped to a dedicated hyperdimensional subspace, realizing a holographic distributed representation of the target features.

[0048] The adaptive quantization module includes a policy selection unit, a binarization unit, and an accumulation unit. The policy selection unit calls the column strength calculation unit to calculate the sum of column strengths of the input fused data and background information. It then determines whether the target is small or large based on the pre-trained target size column strength threshold. Small targets use a quantization-then-accumulation strategy, while large targets use an accumulation-then-quantization strategy. The binarization unit converts the hyper-dimensional projected data into a bipolar vector, which involves sign determination and majority logic operations. The accumulation unit performs accumulation operations on the binary hypervector or floating-point hypervector to obtain the target class vector.

[0049] During the data processing stage, the raw SAR echo data is fused by the data stitching module, and then the column intensity calculation, dynamic segmentation and dimension allocation are completed by the difference mean adaptive segmentation module. It is then mapped to the hyperdimensional space by the hyperdimensional projection module, and finally the target class vector is output by the adaptive quantization module for subsequent Hamming distance calculation with the hyperdimensional vector library.

[0050] In a further embodiment, the data splicing module achieves I / Q dual-channel data fusion through an array column splicing algorithm. The splicing process does not require complex calculations and only completes data recombination through index mapping.

[0051] In a further embodiment, the column strength calculation unit of the difference mean adaptive segmentation module uses a batch summation and mean calculation algorithm to traverse the fused data through a sliding window. The system calculates the mean absolute value of the differences in each column in batches, avoiding the inefficiency caused by element-by-element iteration; the dynamic segmentation unit divides column data through cumulative sums and threshold judgments, and determines the number of segments. The dimensions can be flexibly adjusted through software configuration; the dimension allocation unit adopts a linear allocation rule, allocating the super-dimensional dimensions according to the average proportion of the differences in each segment, ensuring that resources are tilted towards the segments where the target features are concentrated.

[0052] In a further embodiment, the hyperdimensional matrix of the hyperdimensional projection module is a pre-stored hyperdimensional matrix generated during the training phase. The values ​​of the matrix elements are controlled by a random seed, and dynamic updates are supported for different scenarios. The projection operation is implemented through element-wise multiplication and accumulation of vectors and matrices. Parallel computing can be used to optimize and improve processing speed, adapting to the high-efficiency processing requirements of massive echo data.

[0053] In a further embodiment, the strategy selection unit of the adaptive quantization module has a built-in target size judgment threshold, which automatically identifies the target size by statistically summing the column strengths. The strategy of quantizing first and then accumulating uses a sign function to convert the projected data into a binary vector before performing integer accumulation, thus strengthening the signal weight of small targets. The strategy of accumulating first and then quantizing retains the target amplitude information through floating-point accumulation, and then converts it into a binary vector using a sign function, thus maintaining the accuracy of large target discrimination.

[0054] In a further embodiment, after encoding, the vector is compared bit-by-bit with the pre-stored standard class hypervectors of each category in the hyperdimensional vector library, and the Hamming distance between the vectors is calculated. If the Hamming distance between the target class vector and a certain category's standard class hypervector is less than or equal to a threshold, it is determined that a target of that category exists in the input SAR raw echo; if all Hamming distances are greater than the threshold, it is determined that there is no target in the scene. Through this similarity matching process, the final output is a SAR target detection result containing target presence and category information, realizing end-to-end processing from raw echo data to target detection conclusions.

[0055] The logical ideas behind the methods disclosed in the above embodiments can be implemented, in whole or in part, through software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] 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 target screening method based on binary hyperdimensional calculation, characterized in that, Includes the following steps: S1. Obtain the raw SAR echo I / Q dual-channel complex data, and concatenate the real and imaginary matrices to generate a fused data matrix. S2. Based on the pre-stored original echo data of the noisy background, calculate the mean absolute value of the difference between each column in the fused data matrix and the corresponding background column to obtain the column intensity vector; S3. Accumulate all columns of the fused data matrix using a dynamic segmentation algorithm to obtain the cumulative sum of column strengths, and divide the fused data matrix into K segments based on the cumulative sum of column strengths, so that the cumulative sum of column strengths of each fused data sub-matrix tends to be equal. S4. Apply an independent random binary projection submatrix to each segment of the fused data submatrix to perform hyperdimensional projection, and obtain the floating-point projection results of each segment. S5. Determine the target size type based on the cumulative sum of column strengths described in step S3: If the cumulative sum of column strengths is less than a preset threshold, then the floating-point projection results of each segment are first quantized by a sign function and then accumulated. Otherwise, the floating-point projection results of each segment are first accumulated and then quantized by the sign function to finally generate the sample hypervector; S6. Calculate the Hamming distance between the sample hypervector and the pre-stored background reference hypervector. If the Hamming distance is less than or equal to the preset judgment threshold, it is determined that there is no target; otherwise, it is determined that there is a target.

2. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, In step S1, the real and imaginary parts of the original SAR echo I / Q dual-channel complex data are extracted respectively, forming two dimensions. The real-valued matrix is ​​concatenated along the column direction to form a single matrix, generating a matrix of dimension . The fused data matrix.

3. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, In step S2, the pre-stored noise background raw echo data is obtained by multi-frame bit-by-bit accumulation of SAR echo samples in a targetless scene and taking the average value, which is denoted as the background reference matrix. For each column in the fused data matrix, calculate the absolute difference between it and the corresponding column element of the background reference matrix, and average the values ​​of all rows in that column to obtain a column intensity vector of length 2N.

4. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, In step S3, the dynamic segmentation algorithm includes: Calculate the cumulative sum of column strengths and iterate through the range of candidate segment number K. For each candidate segment number K, divide the 2N columns into K segments according to the threshold proportional to the cumulative sum of column strengths, so that the cumulative sum of column strengths in each segment tends to be equal. Calculate the Hamming distance difference between the hypervectors generated by the target sample and the background sample under the current segmentation scheme, select K that maximizes the difference as the optimal number of segments, and record the start and end column indices and segment lengths of each segment.

5. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, Step S4 specifically includes: Each fused data submatrix is ​​independently multiplied by its corresponding random binary projection submatrix. This random binary projection submatrix is ​​pre-generated during the training phase, with its elements independently and identically distributed in {-1, +1}, and has a dimension of 1. ,in The hyperdimensional dimension assigned to the i-th segment; The column number of the i-th segment is the segment length; the projection result is a segment. A 3D floating-point vector.

6. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, Step S5 specifically includes: Compare the cumulative sum of column strengths of the fused data matrix with a preset target size threshold: If the cumulative sum of column strengths is less than a preset threshold, it is determined to be a small target. The projection results of each segment are first converted into a binary matrix by applying a sign function to each element, then accumulated by column to obtain an integer vector, and finally the sign is taken again to generate the final sample hypervector. If the cumulative sum of column strengths is greater than or equal to a preset threshold, it is determined to be a large target. First, the segments are accumulated by floating point to obtain a single vector, and then a sign function is applied uniformly for quantization.

7. The SAR target screening method based on binary hyperdimensional calculation according to claim 1, characterized in that, In step S6, the pre-stored background reference hypervector is generated by performing majority voting on the sample hypervectors obtained after executing steps S1 to S5 on the targetless scene sample; during inference, the Hamming distance between the query sample hypervector and the background reference hypervector is calculated. If the distance is less than or equal to a preset judgment threshold, it is determined that there is a target; otherwise, it is determined to be pure background.

8. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the SAR target screening method based on binary hyperdimensional calculation as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on an electronic device, causes the electronic device to perform the SAR target screening method based on binary hyperdimensional calculation as described in any one of claims 1 to 7.

Images