Weak and small target detection preprocessing method and device based on Gaussian kernel matching

Abstract

The invention relates to the technical field of radar signal processing, and discloses a weak and small target detection preprocessing method and device based on Gaussian kernel matching. The method comprises the steps that radar original echoes are acquired, and an original feature image is generated through two-dimensional discrete Fourier transform and pseudo-color mapping; constructing a gravitational field model on the feature image, and performing superposition calculation on gravitational potential energy and internal stress of a pixel neighborhood to obtain a gravitational field enhanced feature image; constructing a multi-scale Gaussian template and generating an initialized weight matrix; and constructing a multi-scale Gaussian matching module as a first layer of the network, performing feature extraction and channel weighted fusion on the gravitational field enhanced feature image by using a final weight matrix superposed with a preset perturbation parameter as a fixed convolution kernel, and outputting a weighted feature map with physical features. According to the method, speckle noise is suppressed through a physical model, the problem of target scale uncertainty is solved by using Gaussian prior, and the detection performance of a subsequent target detection network on a radar weak and small target is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar signal processing and deep learning, and particularly relates to a preprocessing method for weak and small target detection in a radar range-Doppler map. BACKGROUND

[0002] Radar is an electronic device that detects the position, speed and other characteristics of objects by transmitting and receiving radio waves, and has the advantages of all-weather and all-day detection, and is the core equipment of modern air defense early warning, sea surface monitoring and border security. However, with the wide application of unmanned aerial vehicles and other "low, slow and small" (LSS) targets, radar detection is facing severe challenges. The radar scattering cross section of such targets is extremely small, and the echo energy is weak; at the same time, their flight height is low and speed is slow, and they are easily submerged in complex ground clutter, sea clutter and the inherent coherent speckle noise of the radar system. On the range-Doppler (RD) spectrum or range-azimuth (RA) spectrum, these weak and small targets often only show as isolated pixels or weak spots with extremely low signal-to-noise ratio, resulting in great difficulty in detection.

[0003] Traditional radar target detection mainly relies on the constant false alarm rate (CFAR) algorithm. However, the existing CFAR method has the following significant defects: first, it has poor adaptability to non-uniform clutter and is difficult to accurately describe the complex background, resulting in a high false alarm rate; second, the CFAR algorithm is seriously disturbed by the coherent speckle noise, and the coherent speckle determined by the radar coherent imaging mechanism has the characteristics of multiplicative noise, and the CFAR is difficult to distinguish between isolated strong coherent speckle and real weak and small point targets; finally, when there are multiple targets in the reference window, the background mean value estimation will be pulled up, resulting in missed detection of nearby weak targets.

[0004] In recent years, deep learning algorithms represented by convolutional neural networks and YOLO series have achieved great success in optical image target detection and are gradually introduced into the field of radar signal processing. However, due to the lack of effective signal preprocessing mechanism, the network designed for natural optical images is directly applied to radar spectrum, which has obvious physical mechanism mismatch. First, the radar spectrum lacks the rich texture, color and edge features of the optical image. Under the condition of low signal-to-noise ratio, the features of small weak targets are extremely unobvious. If no targeted enhancement processing is performed before inputting the network, the convolutional layer often has difficulty in extracting effective target features from strong noise and speckle, and even the noise pattern may be learned as target features, resulting in poor model generalization ability. Secondly, the existing deep learning framework usually adopts random initialization (such as Xavier or Kaiming initialization) strategy, assuming that the features are randomly distributed in space. However, the imaging of radar point targets on the spectrum has a clear physical property - its point spread function (PSF) is approximately a Sinc function, and its main lobe presents a peak shape similar to a Gaussian distribution in the two-dimensional plane. The random initialized convolution kernel presents a disordered state at the beginning of training, and has no perception ability for this specific physical form, resulting in the network needing to consume a large number of training iteration times to "blindly search" and fit this Gaussian form. In addition, in actual radar detection, affected by the change of detection distance, motion compensation residual or defocusing effect, the size (i.e. scale) of the target spot on the RD / RA spectrum is not fixed. Some targets show sharp single points (small scale), and some show diffuse patches (large scale). The first layer of the existing general detection network usually uses a single size of convolution kernel, lacks a preprocessing interface that can adapt to the multi-scale characteristics of radar, and is difficult to simultaneously consider the best matching of targets of different scales. In addition, the existing network lacks the ability to adaptively select different scale features, and cannot dynamically adjust the weights of different scale channels according to the characteristics of the input signal, which may easily introduce irrelevant scale noise interference.

[0005] In summary, there is a lack of a comprehensive preprocessing scheme in the prior art that can simultaneously solve the signal-to-noise ratio enhancement under strong speckle interference, the adaptation of the network model to the physical characteristics of the radar point target, and the uncertainty of the target scale. How to construct a high-quality preprocessing mechanism, suppress speckle at the data end using the physical model, and construct a network input interface using multi-scale physical priors to provide high signal-to-noise ratio input features and physical prior parameters for the subsequent detection network is a key technical problem to be solved in the field of radar small weak target detection. The present application aims to solve the above problems. SUMMARY

[0006] Therefore, it is necessary to provide a small weak target detection preprocessing method and device based on Gaussian kernel matching, which can improve the precision detection and anti-interference ability of small weak targets.

[0007] A weak and small target detection preprocessing method based on Gaussian kernel matching, the method comprising:

[0008] Step S100: radar raw echo, global parameter input and radar echo feature mapping; obtain two-dimensional time domain radar raw echo data collected by a radar receiver, and input global configuration parameters; the global configuration parameters include: radar signal processing related parameters, gravity field enhancement algorithm parameters and detection network related parameters; wherein the radar signal processing related parameters include sampling rate, pulse repetition frequency, distance sampling point number, pulse number, speed of light; the gravity field enhancement algorithm parameters include gravity field action coefficient and gravity field action radius; the detection network related parameters include network input image height and width, Gaussian template scale number, Gaussian variance, convolution kernel size and preset perturbation parameter matrix; perform two-dimensional discrete Fourier transform of the radar raw echo data in the distance direction and the Doppler direction according to the radar signal processing related parameters to obtain a frequency domain complex matrix, calculate the logarithmic modulus value of the frequency domain complex matrix to obtain an amplitude spectrum and resample it to a network input size, and convert the amplitude spectrum into a three-channel radar original feature image using a Jet color mapping function;

[0009] Step S200: image feature enhancement based on a gravity field model; obtain the radar original feature image output by step S100, and construct a two-dimensional gravity field model on the radar original feature image according to the gravity field action coefficient and the gravity field action radius input by step S100; calculate the neighborhood gravity potential energy of each pixel in the image and the internal stress of the pixel itself, superimpose the two to obtain an enhanced single-channel intensity value, and convert it into a three-channel gravity field enhanced feature image using a Jet color mapping function;

[0010] Step S300: weight matrix construction based on multi-scale Gaussian prior; according to the Gaussian template scale number, Gaussian variance and convolution kernel size input by step S100, construct a corresponding two-dimensional Gaussian distribution template for each scale; perform coordinate discretization sampling on the template to obtain an original weight matrix, and perform energy normalization processing on the matrix to generate multiple groups of Gaussian initialization weight matrices with different sizes;

[0011] Step S400: detecting the construction, initialization and feature output of the first layer structure of the network; constructing a multi-scale Gaussian matching module containing multiple parallel convolution branches and a subsequent channel weighted fusion module as the first layer structure of the target detection network; using the multiple sets of Gaussian initialization weight matrices generated in step S300 to initialize the weight parameters of each convolution branch; receiving the gravity field enhanced feature image output in step S200 and inputting it into the multi-scale Gaussian matching module to extract multi-scale features through parallel convolution operation, and calculating the importance weight of each scale to perform weighted fusion through the channel weighted fusion module, and outputting the weighted and fused feature map as the input data of the subsequent target detection backbone network.

[0012] Further, the specific process in step S100 includes:

[0013] Step S101, parameter input and analysis; receiving the global configuration parameters input externally and analyzing to obtain the following variables: radar signal processing related parameters (including sampling rate, pulse repetition frequency f r , distance sampling point number N rg , pulse number N pulse ); gravity field enhancement algorithm parameters (including gravity field action coefficient m, gravity field action radius R); detection network related parameters (including network input image height H and width W, Gaussian template scale number K, Gaussian variance {σ1, σ2,..., σ K}, convolution kernel size S k , preset perturbation parameter matrix ΔW (k) );

[0014] Step S102, radar echo acquisition; acquiring the two-dimensional time domain radar original echo signal S(n, m) collected by the radar receiver input externally, where n is the fast time sampling index (0≤n<N rg ), and m is the slow time pulse index (0≤m<N pulse );

[0015] Step S103, two-dimensional transformation; performing two-dimensional discrete Fourier transform (DFT) on S(n, m) in the distance direction and the Doppler direction to obtain a frequency domain complex matrix D(k, l):

[0016]

[0017] wherein D(k, l) represents the complex value of the frequency domain complex matrix at the kth distance frequency unit and the lth Doppler frequency unit; ∑ is a summation symbol, representing the accumulation operation of the values in the specified index range; m is a slow time dimension pulse index variable, taking a value range of 0 to N pulse -1; N pulseThe number of pulses parsed in step S101 represents the total length of the slow time dimension; n is the sampling point index variable of the fast time dimension, and the value range is 0 to N rg -1; N rg The number of distance direction sampling points parsed in step S101 represents the total length of the fast time dimension; S(n, m) is the complex sampling value of the radar two-dimensional time domain echo signal at the nth fast time sampling point and the mth pulse; e is the base of natural logarithm (Euler number); j is the imaginary unit, and satisfies j 2 =-1; π is the circular constant;

[0018] Step S104, modulus calculation and size resampling; Calculate the logarithmic modulus value A(k, l) of the complex matrix:

[0019]

[0020] Wherein, A small constant is used to prevent the singularity of the logarithm, and |·| represents the complex modulus operation;

[0021] Use the nearest neighbor interpolation algorithm to resample the N rg ×N pulse dimensional matrix A into an H×W dimensional matrix A';

[0022] Step S105, normalization; linearly normalize the resampled matrix A' to obtain V(i, j), and the calculation formula is:

[0023]

[0024] Wherein, V(i, j) is the normalized pixel value, and the value range is [0, 1]; (i, j) is the image coordinate index, and satisfies 0≤i

[0025] Step S106, pseudo-color mapping; use the Jet color map mapping function to map the normalized value V(i, j) to the red (R), green (G), and blue (B) three-channel pixel values to generate the radar original feature image I, and the calculation formula is:

[0026] I R (i, j) = clamp (1.5- |4·V(i, j)-3|, 0, 1)

[0027] I G (i, j) = clamp (1.5- |4·V(i, j)-2|, 0, 1)

[0028] IB (i, j) = clamp(1.5 - |4·V(i, j) - 1|, 0, 1)

[0029] Where, I R , I G , I B respectively represent the pixel intensity values of the red F s , green, and blue channels; clamp(x, a, b) is a truncation function defined as: return a when x < a, return b when x > b, otherwise return x; H represents the absolute value operation; 1.5, 4, 3, 2, 1 are all coefficients of the fixed piecewise linear function of the Jet color mapping algorithm.

[0030] Furthermore, the specific process described in step S200 includes:

[0031] Step S201, determining the gravitational field parameters; obtaining the gravitational field action coefficient m and the gravitational field action radius R analyzed in step S101; for any pixel point (i, j) in the radar original feature image I, determining its neighborhood set Ω (i,j) with a radius of R; defining h(i, j) as the grayscale intensity value of the pixel point (i, j), in order to convert the three-channel color information into a single-channel physical scalar field that can characterize the strength of the target energy, the standard brightness weighting formula is used for calculation:

[0032] h(i, j) = 0.299×I R (i, j) + 0.587×I G (i, j) + 0.114×I B (i, j)

[0033] Where, I R (i, j), I G (i, j), I B (i, j) are respectively the pixel values of the red, green, and blue channels of the radar original feature image I(i, j) at the coordinate (i, j) generated in step S106; the coefficients 0.299, 0.587, and 0.114 are standard grayscale conversion weights that conform to human visual perception and can retain the energy distribution characteristics in the Jet pseudocolor mapping to the greatest extent;

[0034] Step S202, calculating the neighborhood gravitational potential energy; traversing all pixel points (u, v) in the neighborhood set Ω (i，j) and calculating the total gravitational contribution of them to the central pixel point (i, j) Its calculation formula is:

[0035]

[0036] Where, represents the sum of the neighborhood attractive potential energy at pixel point (i, j); Ω (i,j) represents the neighborhood pixel set with (i, j) as the center and radius R; m is the attractive field action coefficient, used to adjust the strength of the attractive enhancement; h(i, j) is the original gray intensity value of the center pixel point (i, j); h(u, v) is the original gray intensity value of a pixel point (u, v) in the neighborhood; d is the Euclidean distance between pixel points (i, j) and (u, v), and the calculation formula is ∑ is a summation symbol, representing the cumulative operation on the values in the specified index range;

[0037] Step S203, calculating internal stress; using the intensity of the pixel itself to calculate the internal stress F In (i, j), and the calculation formula is:

[0038] F In (i, j) = m·[h(i, j)] 2

[0039] wherein, F In (i, j) represents the internal stress at pixel point (i, j), used to maintain the target edge sharpness; m is the attractive field action coefficient; h(i, j) is the original gray intensity value of pixel point (i, j);

[0040] Step S204, superimposed to generate an enhanced image; calculating the attractive field enhanced feature image I gray corresponding to the new intensity value of the pixel point in the middle, and the calculation formula is:

[0041]

[0042] wherein, I gray (i, j) is the pixel value of the enhanced image at coordinate (i, j); is the neighborhood attractive potential energy calculated in step S202; F In (i, j) is the internal stress calculated in step S203.

[0043] I gray is linearly normalized, and then converted into a three-channel attractive field enhanced feature image I ò using the Jet color map mapping function described in step S106 again, the dimension of the image is HxWx3, and it is used as the input data of the target detection neural network in step S400.

[0044] Further, the specific process in step S300 includes:

[0045] Step S301, parameter analysis; the Gaussian template scale number K and the convolution kernel size S k(S k K sets of Gaussian variance parameters {σ1, σ2,..., σK} are determined, where K is an odd number. K};

[0046] Step S302, discretization sampling; for the kth scale (1≤k≤K), traverse the row index r and column index c (0≤r, c<S k ) of the convolution kernel matrix, and calculate the coordinates (x, y) relative to the center of the convolution kernel:

[0047]

[0048] where x and y are the coordinates relative to the center of the convolution kernel; S k is the size of the convolution kernel; r and c are the row index and column index of the convolution kernel matrix; represents the floor operation;

[0049] Substitute (x, y) into the two-dimensional Gaussian formula to obtain the unnormalized discrete weight element The calculation formula is:

[0050]

[0051] where, represents the original weight value of the Gaussian template of the kth scale at the matrix coordinates (r, c); exp is the exponential function with the natural constant e as the base; x, y are the horizontal and vertical distances relative to the center of the convolution kernel; σ k is the kth set of Gaussian variance parameters, used to control the distribution width of the Gaussian function;

[0052] Step S303, energy normalization; calculate the element sum Sum of the discrete weight matrix k , and perform normalization processing on the matrix to obtain the initialized weight matrix The calculation formula is:

[0053]

[0054] where Sum k is the sum of all elements of the kth unnormalized weight matrix; is the normalized initialized weight value, satisfying the sum of the matrix elements is 1; is the discrete weight matrix calculated in step S302.

[0055] Further, the specific process in step S400 includes:

[0056] Step S401, constructing a parallel structure; a multi-scale Gaussian matching module is constructed before the backbone network of the target detection network, which contains K parallel convolution branches; the convolution kernel size of the kth branch is configured as S k , and the zero padding parameter P is set k , and the calculation formula is:

[0057]

[0058] Wherein, P k represents the number of layers of the kth branch filled with 0 at the edge of the input image, which is used to ensure that the spatial size of the feature map output by all branches remains HxW; S k is the convolution kernel size; represents the floor operation;

[0059] Step S402, weight synthesis and assignment; the preset perturbation parameter matrix parsed in step S101 is superimposed with the initialization weight matrix of the corresponding scale generated in step S303, and the final convolution kernel weight is calculated:

[0060]

[0061] Wherein, is the fixed Gaussian prior initial weight generated in step S300; ΔW (k) is the preset perturbation parameter matrix parsed in step S101, which is used to fit the shape deviation of the real radar target relative to the standard Gaussian model. After the final convolution kernel weight is expanded in channel number, it is assigned to the weight parameter of the corresponding convolution branch as a fixed convolution kernel parameter;

[0062] Step S403, feature extraction and splicing; the gravity field enhanced feature image I ò is input to K parallel convolution branches at the same time, and the gravity field enhanced feature image is convolved using the Gaussian initialization weight matrix assigned in step S402, to obtain K groups of feature maps F1, F2,..., F K of different scales; the K groups of feature maps are spliced in the channel dimension to form a multi-scale feature tensor F concat ;

[0063] Step S404, adaptive scale weighting; a channel weighting fusion module is connected after the spliced multi-scale feature tensor; first, the global average pooling is performed on F concat to obtain a channel descriptor vector z, and the calculation formula of the cth element of z is:

[0064]

[0065] Wherein, z cglobal statistics feature of the cth channel, 1≤c≤C total , dimension feature tensor F concat , total channel number C total , value K; H, W are the height and width of the feature map; F concat (i, j, c) is the value of the feature tensor at coordinate (i, j) and channel c;

[0066] Then the weight vector a is generated by using two fully connected layers and Sigmoid activation function, and the calculation formula is:

[0067] a = sigma (W2delta (W1z))

[0068] wherein, is the channel attention weight vector of the output; sigma is the Sigmoid function; delta is the ReLU activation function; is the channel descriptor vector mentioned above; is the weight matrix of the first dimension reduction fully connected layer, and r is the dimension reduction ratio; is the weight matrix of the second dimension increasing fully connected layer;

[0069] Finally, the weight vector a is multiplied with the feature tensor F concat channel by channel, and the weighted fused feature map is output.

[0070] A weak small target detection preprocessing device based on Gaussian kernel matching, the device comprises:

[0071] The radar original echo, global configuration parameter input and radar echo feature mapping module are used for acquiring two-dimensional time domain radar original echo data collected by a radar receiver, and inputting global configuration parameters; the global configuration parameters include radar signal processing related parameters, gravity field enhancement algorithm parameters and detection network related parameters; and are also used for performing two-dimensional discrete Fourier transform on the radar original echo data in the distance direction and the Doppler direction to obtain a frequency domain complex matrix according to the radar signal processing related parameters, calculating the logarithmic modulus value of the frequency domain complex matrix to obtain an amplitude spectrum and resampling to a network input size, and converting the amplitude spectrum into a three-channel radar original feature image by using a Jet color mapping function;

[0072] The image feature enhancement module based on the gravity field model is used for acquiring the radar original feature image, and constructing a two-dimensional gravity field model on the radar original feature image according to the gravity field action coefficient and the gravity field action radius in the gravity field enhancement algorithm parameters; the neighborhood gravity potential energy of each pixel point in the image and the internal stress of the pixel itself are calculated, the two are superimposed to obtain an enhanced single-channel intensity value, and the Jet color mapping function is used to convert the single-channel intensity value into a three-channel gravity field enhanced feature image;

[0073] A weight matrix construction module based on a multi-scale Gaussian prior, configured to construct a corresponding two-dimensional Gaussian distribution template for each scale according to a Gaussian template scale number, a Gaussian variance and a convolution kernel size in the detection network related parameters; perform coordinate discretization sampling on the template to obtain an original weight matrix, and perform energy normalization processing on the matrix to generate a plurality of groups of Gaussian initialization weight matrices with different sizes;

[0074] A detection network first layer structure construction, initialization and feature output module, configured to construct a multi-scale Gaussian matching module containing a plurality of parallel convolution branches and a channel weighted fusion module, and use the multi-scale Gaussian matching module as the first layer structure of the target detection network; initialize the weight parameters of each convolution branch by using the final convolution kernel weight generated by superimposing the plurality of groups of Gaussian initialization weight matrices and the preset perturbation parameter matrix; receive the gravity field enhanced feature image, input the gravity field enhanced feature image into the multi-scale Gaussian matching module, extract multi-scale features through parallel convolution operation, calculate the importance weight of each scale through the channel weighted fusion module to perform weighted fusion, and output the weighted fused feature map as the input data of the subsequent target detection backbone network.

[0075] Compared with the prior art, the present application has the following advantages: 1) the present application can provide high-quality enhanced features and physical perception ability for the subsequent target detection network in a low signal-to-noise ratio radar complex electromagnetic environment through the combination of the gravity field physical model and the deep learning Gaussian prior, thereby assisting in realizing accurate detection and anti-interference of weak and small targets; 2) the present application takes the gravity field model as the preprocessing core; utilizes the spatial gravity interaction between pixels, calculates the neighborhood gravity potential energy and internal stress, and realizes effective suppression of random speckle noise and aggregation of target energy at the physical level, thereby significantly improving the signal-to-noise ratio at the data input end and providing purer input data for the deep network; 3) the present application constructs a multi-scale Gaussian matching preprocessing interface based on physical prior, matches radar targets with different distance resolutions or defocusing degrees through convolution kernels with different sizes, and introduces an adaptive scale weighting mechanism, so that the preprocessing module can dynamically adjust the importance of each scale according to the input signal characteristics, thereby effectively alleviating the scale uncertainty problem of radar targets caused by motion compensation error or distance movement at the data input stage; 4) the present application uses a parameter fine-tuning strategy based on physical prior, takes the Gaussian kernel as the initialization reference point of the preprocessing module, allows the network to learn perturbation parameters, effectively avoids the randomness and uncertainty of parameter optimization of the network at the initial training stage, and gives the model input end the flexibility to adapt to real non-ideal target shapes (such as sidelobe distortion). BRIEF DESCRIPTION OF DRAWINGS

[0076] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. It is readily apparent to one skilled in the art that the accompanying drawings typically illustrate only some embodiments of the application and therefore should not be considered as limiting the scope of the application, as there can be numerous variations and modifications.

[0077] Figure 1 A flow chart of a weak and small target detection preprocessing method based on Gaussian kernel matching is shown according to an embodiment of the application;

[0078] Figure 2 A principle and effect comparison schematic diagram of image feature enhancement based on a gravitational field model is shown according to an embodiment of the application.

[0079] Figure 3 An internal processing flow and feature fusion mechanism schematic diagram of a multi-scale Gaussian matching module is shown according to an embodiment of the application.

[0080] Figure 4 An overall architecture and data and parameter flow schematic diagram of a weak and small target detection preprocessing device is shown according to an embodiment of the application.

[0081] Figure 5 A structural block diagram of a weak and small target detection preprocessing device based on Gaussian kernel matching is shown according to an embodiment of the application. DETAILED DESCRIPTION

[0082] It should be noted that the embodiments and features of the embodiments in the present application can be combined with each other without conflict. The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The description of the at least one exemplary embodiment is actually only illustrative, but not intended to limit the application and use of the present application in any way. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0083] It should be noted that the terms used herein are only intended to describe specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of the features, steps, operations, devices, components and / or combinations thereof.

[0084] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application and not to limit the present application.

[0085] The relative arrangement of components and steps, numerical expressions, and values set forth in these embodiments do not limit the scope of the present application unless otherwise specifically stated. It should be understood that the dimensions of the various parts shown in the drawings are not drawn to scale for the sake of convenience in description. Techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail, but should be considered as part of the authorized description where appropriate. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments can have different values. It should be noted that similar reference numbers and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0086] As shown in Figure 1 The present application provides a weak target detection preprocessing method based on Gaussian kernel matching, which comprises:

[0087] Step S100: radar raw echo, global parameter input and radar echo feature mapping; obtain two-dimensional time domain radar raw echo data collected by a radar receiver, and input global configuration parameters; the global configuration parameters include radar signal processing related parameters, gravity field enhancement algorithm parameters and detection network related parameters; wherein the radar signal processing related parameters include sampling rate, pulse repetition frequency, distance sampling point number, pulse number, and light speed; the gravity field enhancement algorithm parameters include gravity field action coefficient and gravity field action radius; the detection network related parameters include network input image height and width, Gaussian template scale number, Gaussian variance, convolution kernel size, and pre-set perturbation parameter matrix; perform two-dimensional discrete Fourier transform of the radar raw echo data in the distance and Doppler directions to obtain a frequency domain complex matrix according to the radar signal processing related parameters, calculate the logarithmic modulus value of the frequency domain complex matrix to obtain an amplitude spectrum and resample it to a network input size, and convert the amplitude spectrum into a three-channel radar raw feature image using a Jet color mapping function.

[0088] Step S200: image feature enhancement based on gravitational field model; obtain the radar original feature image output by step S100, and construct a two-dimensional gravitational field model on the radar original feature image according to the gravitational field action coefficient and the gravitational field action radius input by step S100; calculate the neighborhood gravitational potential energy of each pixel point in the image and the internal stress of the pixel itself, superimpose the two to obtain an enhanced single-channel intensity value, and convert it into a three-channel gravitational field enhanced feature image by using a Jet color mapping function.

[0089] Step S300: weight matrix construction based on multi-scale Gaussian prior; according to the Gaussian template scale number, Gaussian variance and convolution kernel size input by step S100, a corresponding two-dimensional Gaussian distribution template is constructed for each scale; the template is sampled by coordinate discretization to obtain an original weight matrix, and the matrix is energy-normalized to generate a plurality of Gaussian initialization weight matrices of different sizes.

[0090] Step S400: construction, initialization and feature output of the first layer structure of the detection network; a multi-scale Gaussian matching module containing a plurality of parallel convolution branches and a channel weighting fusion module is constructed, which is used as the first layer structure of the target detection network; the plurality of Gaussian initialization weight matrices generated by step S300 are used to initialize the weight parameters of each convolution branch; the gravitational field enhanced feature image output by step S200 is received and input into the multi-scale Gaussian matching module, multi-scale features are extracted through parallel convolution operation, and the importance weight of each scale is calculated through the channel weighting fusion module to perform weighted fusion, and the weighted and fused feature map is output as the input data of the subsequent target detection backbone network.

[0091] The present application can provide high-quality enhanced features and physical perception ability for subsequent target detection network in a low signal-to-noise ratio radar complex electromagnetic environment by combining the gravitational field physical model with the deep learning Gaussian prior, thereby assisting in realizing accurate detection and anti-interference of weak and small targets; the gravitational field model is used as the preprocessing core; the spatial gravitational interaction between pixels is used to calculate the neighborhood gravitational potential energy and internal stress, thereby effectively suppressing random speckle noise and aggregating target energy at the physical level, thereby significantly improving the signal-to-noise ratio at the data input end and providing purer input data for the deep network; a multi-scale Gaussian matching preprocessing interface based on physical prior is constructed, different size convolution kernels are used to match radar targets with different distance resolutions or defocus degrees, and an adaptive scale weighting mechanism is introduced, so that the preprocessing module can dynamically adjust the importance of each scale according to the input signal characteristics, thereby effectively relieving the scale uncertainty problem of radar targets caused by motion compensation error or distance movement at the data input stage; a parameter fine-tuning strategy based on physical prior is used, and the Gaussian kernel is used as the initialization reference point of the preprocessing module, allowing the network to learn perturbation parameters, which not only effectively avoids the randomness and uncertainty of parameter optimization of the network at the initial training stage, but also gives the model input end the flexibility to adapt to real non-ideal target morphology (such as sidelobe distortion).

[0092] According to an embodiment of the present application, in step S100 of the present application, the radar original echo, global parameter input and radar echo feature mapping include:

[0093] Step S101, parameter input and analysis; receive the externally input global configuration parameters and analyze to obtain the following variables: radar signal processing related parameters (including sampling rate F s , pulse repetition frequency f r , distance sampling point number N rg , pulse number N pulse ); gravitational field enhancement algorithm parameters (including gravitational field action coefficient m, gravitational field action radius R); detection network related parameters (including network input image height H and width W, Gaussian template scale number K, Gaussian variance {σ1, σ2,..., σ K}, convolution kernel size S k , preset perturbation parameter matrix ΔW (k) );

[0094] Step S102, radar echo acquisition; acquire the two-dimensional time domain radar original echo signal S(n, m) collected by the radar receiver externally input, wherein n is the fast time sampling index (0≤n<N rg ), m is the slow time pulse index (0≤m<N pulse );

[0095] Step S103, two-dimensional transform; a two-dimensional discrete Fourier transform (DFT) is performed on S(n, m) in the range and Doppler direction to obtain a frequency domain complex matrix D(k, l):

[0096]

[0097] wherein D(k, l) represents a complex value of the frequency domain complex matrix at the kth range frequency unit and the lth Doppler frequency unit; ∑ is a summation symbol, representing an accumulation operation on the values in the specified index range; m is a pulse index variable in the slow time dimension, and the value range is 0 to N pulse -1; N pulse is the number of pulses parsed in step S101, representing the total length of the slow time dimension; n is a sampling point index variable in the fast time dimension, and the value range is 0 to N rg -1; N rg is the number of range sampling points parsed in step S101, representing the total length of the fast time dimension; S(n, m) is a complex sampling value of the radar two-dimensional time domain echo signal at the nth fast time sampling point and the mth pulse; e is the base of natural logarithm (Euler number); j is an imaginary unit, and j 2 = -1; π is a circular constant;

[0098] Step S104, modulus calculation and size resampling; the logarithmic modulus A(k, l) of the complex matrix is calculated:

[0099]

[0100] wherein, is a small constant to prevent logarithmic singularity, and |·| represents a complex modulus operation;

[0101] The N rg × N pulse dimensional matrix A is resampled into an H × W dimensional matrix A'; this method directly selects the pixel value of the nearest integer coordinate point in the source matrix as the sampling result by calculating the relative position of the target pixel in the source matrix, and the calculation formula is:

[0102]

[0103] wherein A'(i, j) is a pixel value of the i th row and the j th column after resampling; A(·, ·) is a pixel value in the original logarithmic modulus matrix; round(·) represents a rounding integer operation; N rg / H and N pulse / W are scaling proportion factors in the range direction and the pulse direction, respectively;

[0104] Step S105, normalization; linear normalization is performed on the resampled matrix A' to obtain V(i, j), and the calculation formula is as follows:

[0105]

[0106] wherein V(i, j) is a normalized pixel value, the value range is [0, 1]; (i, j) is an image coordinate index, satisfying 0≤i<H, 0≤j<W; A'(i, j) is the value of the resampled logarithmic modulus matrix in the i-th row and the j-th column; min(A') is the minimum value of all elements in the matrix A'; max(A') is the maximum value of all elements in the matrix A';

[0107] Step S106, pseudo-color mapping; the normalized value V(i, j) is mapped to red (R), green (G), and blue (B) three-channel pixel values by using a Jet color map mapping function, to generate a radar original feature image I, and the calculation formula is as follows:

[0108] I R (i, j) = clamp (1.5 - |4·V(i, j) - 3|, 0, 1)

[0109] I G (i, j) = clamp (1.5 - |4·V(i, j) - 2|, 0, 1)

[0110] I B (i, j) = clamp (1.5 - |4·V(i, j) - 1|, 0, 1)

[0111] wherein I R , I G , I B respectively represent the pixel intensity values of the red, green, and blue channels; clamp (x, a, b) is a clipping function, defined as: returning a when x < a, returning b when x > b, and otherwise returning x; |·| represents an absolute value operation; 1.5, 4, 3, 2, and 1 are fixed piecewise linear function coefficients of the Jet color mapping algorithm.

[0112] According to an embodiment of the present application, in step S200 of the present application, the image feature enhancement based on the gravity field model comprises:

[0113] Step S201, determining the gravity field parameters; obtaining the gravity field action coefficient m and the gravity field action radius R analyzed in step S101; for any pixel point (i, j) in the radar original feature image I, determining a neighborhood set Ω (i，j); define h(i, j) as the gray intensity value of pixel point (i, j), in order to convert three channel color information into a single channel physical scalar field capable of representing the energy intensity of the target, the standard brightness weighted formula is used for calculation:

[0114] h(i, j) = 0.299 * I R (i, j) + 0.587 * I G (i, j) + 0.114 * I B (i, j)

[0115] Wherein, I R (i, j), I G (i, j), I B (i, j) are respectively the red, green and blue channel pixel values of the radar original feature image I(i, j) generated in step S106 at coordinates (i, j); The coefficients 0.299, 0.587 and 0.114 are standard gray scale conversion weights consistent with human eye visual perception, which can maximize the energy distribution characteristics in Jet pseudo-color mapping;

[0116] Step S202, calculate the neighborhood attractive potential; traverse all pixel points (u, v) in the neighborhood set Ω (i,j) , calculate the total contribution of the attractive force to the center pixel point (i, j) The calculation formula is:

[0117]

[0118] Wherein, Indicates the total neighborhood attractive potential of pixel point (i, j); Ω (i，j) Indicates the neighborhood pixel set with (i, j) as the center and radius R; m is the attractive field action coefficient, used to adjust the intensity of attractive force enhancement; h(i,j) is the original gray intensity value of the center pixel point (i, j); h(u, v) is the original gray intensity value of a pixel point (u, v) in the neighborhood; d is the Euclidean distance between pixel points (i, j) and (u, v), and the calculation formula is ∑ is the summation symbol, indicating the cumulative operation on the values in the specified index range;

[0119] Step S203, calculate the internal stress; calculate the internal stress F In (i, j) by using the intensity of the pixel itself, and the calculation formula is:

[0120] F In (i, j) = m * [h(i, j)] 2

[0121] Wherein, F In(i, j) represents the internal stress at pixel point (i, j) for maintaining the target edge sharpness; m is a coefficient of the gravitational field effect; h(i, j) is the original gray intensity value of pixel point (i, j);

[0122] In step S204, the enhanced image is generated by superposition; the gravitational field enhanced feature image I gray is calculated

[0123]

[0124] wherein, I gray (i, j) is the pixel value of the enhanced image at coordinate (i, j); is the neighborhood gravitational potential energy calculated in step S202; F In (i, j) is the internal stress calculated in step S203.

[0125] After linear normalization processing is performed on I gray , the Jet color map mapping function described in step S106 is used again to convert it into a three-channel gravitational field enhanced feature image I ò , which has a dimension of HxWx3 and is used as the input data of the target detection neural network in step S400.

[0126] As shown in Figure 2 , the specific working principle of the gravitational field enhancement module and the effect comparison before and after enhancement are shown. The left side of the figure is a local schematic diagram of the original radar feature image, wherein the pixels in the dashed box represent the center pixel to be processed (the intensity is denoted as h c ), and the pixels in the blue dashed box represent the neighborhood pixels (the intensity is denoted as h r ). According to the gravitational field model proposed in the present application, there is a mutual interaction similar to the universal gravitation between the pixels (as shown by the blue arrows in the figure). The gravitational contribution of the neighborhood pixels to the center pixel is proportional to the product of the intensities of the two and inversely proportional to the square of the distance d (i.e., G∝h c ·h r / d 2 ). The right side of the figure is the image after gravitational field enhancement. By superimposing the neighborhood gravitational potential energy and the internal stress F In of the pixel itself (i.e., ), the energy of the target is significantly aggregated and enhanced. As can be seen from the comparison between the left and right images, the brightness of the center pixel after enhancement (in the red box on the right) is significantly improved, while the intensity of the surrounding isolated background clutter is not amplified to the same degree due to the lack of “gravitational support” of the neighborhood pixels, thereby effectively enlarging the contrast between the target and the background and achieving the improvement of the signal-to-noise ratio.

[0127] According to an embodiment of the present invention, in step S300, the construction of the weight matrix based on multi-scale Gaussian prior includes:

[0128] Step S301, parameter parsing; Step S101 parses the Gaussian template scale number K and convolution kernel size S among the relevant parameters of the detection network. k (S k (where the number is odd) and K groups of Gaussian variance parameters {σ1, σ2, ..., σ... K};

[0129] Step S302, Discretize sampling; for the k-th scale (1≤k≤K), traverse the row index r and column index c (0≤r, c) of the convolution kernel matrix. k ), calculate its coordinates (x, y) relative to the center of the convolution kernel:

[0130]

[0131] Where x and y are coordinates relative to the center of the convolution kernel; S k is the kernel size; r and c are the row and column indices of the kernel matrix; This indicates the floor function;

[0132] Substituting (x,y) into the two-dimensional Gaussian formula yields the unnormalized discrete weight elements. The calculation formula is as follows:

[0133]

[0134] in, σ represents the original weight values ​​of the Gaussian template at the k-th scale at matrix coordinates (r, c); exp is an exponential function with the natural constant e as the base; x and y are the horizontal and vertical distances relative to the center of the convolution kernel; σ k Here is the Gaussian variance parameter for the k-th group, used to control the distribution width of the Gaussian function;

[0135] Step S303, energy normalization; calculate the discrete weight matrix. Sum of elements k The matrix is ​​then normalized to obtain the initial weight matrix. The calculation formula is as follows:

[0136]

[0137] Among them, Sum k It is the sum of all elements of the k-th unnormalized weight matrix; The normalized initial weight values ​​satisfy the condition that the sum of the matrix elements is 1; ​The discrete weight matrix is ​​calculated in step S302.

[0138] According to one embodiment of the present invention, such as Figure 4 As shown, in step S400 of the present invention, the construction, initialization, and feature output of the first layer structure of the detection network include:

[0139] Step S401: Construct a parallel structure; build a multi-scale Gaussian matching module before the backbone network of the object detection network. This module contains K parallel convolutional branches; the kernel size of the k-th branch is configured as S. k And set the zero-fill parameter P k The calculation formula is as follows:

[0140]

[0141] Among them, P k This indicates the number of layers where the k-th branch pads the edges of the input image with zeros, used to ensure that the feature map space size of all branch outputs remains H×W; S k The kernel size; This indicates the floor function;

[0142] Step S402, Weight synthesis and assignment; obtain the preset perturbation parameter matrix parsed in step S101, and superimpose it with the initial weight matrix of the corresponding scale generated in step S303 to calculate the final convolution kernel weights:

[0143]

[0144] in, The fixed Gaussian prior initial weights generated in step S300; ΔW (k) The preset perturbation parameter matrix, analyzed in step S101, is used to fit the shape deviation of the real radar target relative to the standard Gaussian model. The final convolution kernel weights are expanded by the number of channels and then assigned to the weight parameters of the corresponding convolution branches, serving as fixed convolution kernel parameters.

[0145] Step S403, Feature extraction and stitching; the gravitational field enhancement feature image I ò Simultaneously, the input is fed into K parallel convolution branches, and the gravitational field enhancement feature image is convolved using the Gaussian initialization weight matrix assigned in step S402 to obtain K sets of feature maps F1, F2, ..., Fn at different scales. K The K sets of feature maps are concatenated along the channel dimension to form a multi-scale feature tensor F. concat ;

[0146] The specific execution process for this step is as follows:

[0147] First, the gravitational field enhancement feature image output from step S200 is received as network input data. The system simultaneously and in parallel inputs this input image into the K convolutional branches constructed in step S401.

[0148] Next, a physical prior-based convolution operation is performed within each convolutional branch. Taking the k-th branch as an example, this branch has already been assigned a specific Gaussian initialization weight matrix (i.e., the convolution kernel weight W of this branch) in step S402. k In this branch, the system performs a convolution operation between the input image and the Gaussian-initialized weight matrix:

[0149] 1) Sliding window operation: The convolution kernel slides across the input image, and the step size is usually set to 1.

[0150] 2) Convolution calculation: At each sliding position, the elements of the Gaussian initialization weight matrix are multiplied element-wise with the pixel values ​​within the corresponding local window of the image, and then summed. This process utilizes the spatial distribution characteristics of Gaussian weights to weighted aggregate the energy of local regions of the image, thereby extracting the target feature response that matches the Gaussian template at the current scale.

[0151] 3) Scale preservation, since the kernel size S has been preserved in step S401 k The corresponding zero-fill parameter P was configured. k The feature map output by the above convolution operation remains unchanged in spatial size, still being H×W.

[0152] After the above parallel processing, K branches are taken and K sets of single-channel feature maps representing different Gaussian scale responses are output.

[0153] Finally, the system concatenates these K sets of feature maps along the channel dimension. The concatenated result forms a multi-scale feature tensor with dimensions H×W×K. This tensor simultaneously contains the physical morphological features of the input signal at different spatial scales, providing a rich data foundation for subsequent adaptive weighted fusion.

[0154] Step S404, Adaptive Scale Weighting: A channel-weighted fusion module is then connected after the concatenated multi-scale feature tensor; firstly, F... concst Global average pooling is used to obtain the channel descriptor vector z, and the formula for calculating the c-th element is:

[0155]

[0156] Where zc is the global statistical feature of the c-th channel, 1≤c≤C total Scale feature tensor F concat Total number of channels C totalThe value is K; H and W are the height and width of the feature map; F concat (i, j, c) represents the values ​​of the feature tensor at coordinates (i, j) and channel c;

[0157] Next, two fully connected layers and a sigmoid activation function are used to generate a weight vector α, which is calculated as follows:

[0158] α=σ(W2δ(W1z))

[0159] in, σ represents the output channel attention weight vector; σ is the Sigmoid function; δ is the ReLU activation function; The above is the channel descriptor vector; Here is the weight matrix of the first dimension-reduced fully connected layer, and r is the dimension reduction ratio; This is the weight matrix for the second fully connected layer with increased dimensionality.

[0160] Finally, the weight vector α and the feature tensor F are compared. concat Perform channel-by-channel multiplication and output the weighted fused feature map.

[0161] like Figure 3 As shown, this process demonstrates how the multi-scale Gaussian matching module transforms physical priors into network features. First, the gravitational field enhancement feature image output from step S200 is fed into parallel convolutional branches within the module. To address the different morphologies of radar targets in the range-Doppler spectrum due to range resolution differences or motion compensation residuals, each branch is pre-programmed with Gaussian initialization kernels of different scales (3×3, 5×5, and 7×7 matrices as shown in the grid in the figure), specifically designed to capture the features of sharp point targets, standard Gaussian-like targets, and diffuse, out-of-focus targets, respectively. After processing by these convolutional kernels with physical morphology priors, the input image generates multiple sets of single-channel feature maps, which are then stacked and stitched along the channel dimension to form a multi-scale feature tensor containing rich scale information (as shown in the cube in the figure). Subsequently, to achieve feature optimization, this feature tensor undergoes global average pooling and fully connected layer processing, and the network adaptively learns to generate a channel attention weight vector α. Finally, the weight vector is applied to the feature tensor using a multiplication operation (the "X" symbol in the figure) to reorganize the feature channels at different scales, thereby dynamically highlighting the scale features that best match the current echo state and outputting the final weighted fusion feature map for use by the subsequent detection backbone network.

[0162] Compared with existing technologies, this application has the following significant advantages: 1) In complex electromagnetic environments with low signal-to-noise ratio radar, this invention can provide high-quality enhanced features and physical perception capabilities for subsequent target detection networks by combining a gravitational field physical model with deep learning Gaussian priors, thereby assisting in the accurate detection and anti-interference of weak targets; 2) This invention uses a gravitational field model as the core of preprocessing; by utilizing the spatial gravitational interaction between pixels and calculating the neighborhood gravitational potential energy and internal stress, it achieves effective suppression of random speckle noise and aggregation of target energy at the physical level, thereby significantly improving the signal-to-noise ratio at the data input end and providing cleaner input data for deep networks; 3) This invention constructs a physical prior-based... The multi-scale Gaussian matching preprocessing interface of the experiment matches radar targets with different range resolutions or defocusing degrees through convolution kernels of different sizes, and introduces an adaptive scale weighting mechanism, which enables the preprocessing module to dynamically adjust the importance of each scale according to the characteristics of the input signal. This effectively alleviates the scale uncertainty problem caused by the motion compensation error or distance movement of the radar target during the data input stage; 4) The present invention utilizes a parameter fine-tuning strategy based on physical priors, using the Gaussian kernel as the initialization reference point of the preprocessing module, allowing the network to learn perturbation parameters. This effectively avoids the randomness and uncertainty of parameter optimization in the early stage of network training, and gives the model input end the flexibility to adapt to real non-ideal target shapes (such as sidelobe distortion).

[0163] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0164] like Figure 4As shown, the overall architecture of the device described in this invention is characterized by the parallel processing and final convergence of data streams and parameter streams. The system first receives raw echo data collected by the radar receiver and externally input global configuration parameters. After parsing, the global parameters control the operational logic of subsequent modules. In the data path, the raw radar echo first undergoes time-frequency transformation and pseudo-color mapping through the feature mapping module to generate the original radar feature image. It then enters the gravity field enhancement module, which, under the control of the gravity field parameters, uses a physical model to suppress background clutter and enhance target energy, outputting a high signal-to-noise ratio gravity field enhanced feature image. Simultaneously, in the parameter path, the detection network's relevant parameters drive the weight matrix construction module to generate multiple sets of Gaussian initialization weight matrices containing physical priors. Finally, the enhanced image data and the generated Gaussian weight parameters converge at the first-layer structure module of the detection network. The data is input into a multi-scale matching structure initialized by the weight parameters. After convolutional extraction and adaptive weighted fusion, a feature map with physical perception capabilities is output as the input to the subsequent target detection backbone network.

[0165] In one embodiment, a preprocessing device for weak target detection based on Gaussian kernel matching is provided, comprising: a radar raw echo, a global configuration parameter input and radar echo feature mapping module 502, an image feature enhancement module based on a gravitational field model 504, a weight matrix construction module based on multi-scale Gaussian priors 506, and a module for constructing, initializing, and outputting features of the first layer structure of the detection network 508, wherein:

[0166] The radar raw echo, global configuration parameter input, and radar echo feature mapping module 502 are used to acquire two-dimensional time-domain radar raw echo data collected by the radar receiver and input global configuration parameters. The global configuration parameters include radar signal processing related parameters, gravitational field enhancement algorithm parameters, and detection network related parameters. It is also used to perform two-dimensional discrete Fourier transforms in the range and Doppler directions on the radar raw echo data according to the radar signal processing related parameters to obtain a frequency domain complex matrix, calculate the logarithmic modulus of the frequency domain complex matrix to obtain the amplitude spectrum, resample it to the network input size, and use the Jet color mapping function to convert the amplitude spectrum into a three-channel radar raw feature image.

[0167] The image feature enhancement module 504 based on the gravitational field model is used to acquire the original radar feature image and construct a two-dimensional gravitational field model on the original radar feature image according to the gravitational field action coefficient and gravitational field action radius in the gravitational field enhancement algorithm parameters; calculate the neighborhood gravitational potential energy and the internal stress of the pixel itself, superimpose the two to obtain the enhanced single-channel intensity value, and use the Jet color mapping function to convert it into a three-channel gravitational field enhanced feature image;

[0168] The multi-scale Gaussian prior-based weight matrix construction module 506 is used to construct a corresponding two-dimensional Gaussian distribution template for each scale according to the number of Gaussian template scales, Gaussian variance, and convolution kernel size in the relevant parameters of the detection network; to perform coordinate discretization sampling on the template to obtain the original weight matrix; and to perform energy normalization processing on the matrix to generate multiple sets of Gaussian initialization weight matrices of different sizes.

[0169] The detection network first-layer structure construction, initialization, and feature output module 508 is used to construct a multi-scale Gaussian matching module containing multiple parallel convolutional branches and a subsequent channel weighted fusion module, which serves as the first-layer structure of the target detection network. The module initializes the weight parameters of each convolutional branch using the multiple sets of Gaussian initialization weight matrices. It receives the gravitational field enhancement feature image, inputs it into the multi-scale Gaussian matching module, extracts multi-scale features through parallel convolution operations, and calculates the importance weights of each scale through the channel weighted fusion module for weighted fusion. The weighted fused feature map is then output as input data for the subsequent target detection backbone network.

[0170] like Figure 5 As shown, this embodiment also provides a preprocessing device for weak target detection based on Gaussian kernel matching. This device mainly consists of four cascaded core processing modules:

[0171] Radar raw echo, global configuration parameter input and radar echo feature mapping module 502: As the input interface of the device, it is responsible for receiving the raw echo and parsing the global parameters, and converting the echo into a visualized raw feature image through the internal FFT unit and pseudo-color mapping unit.

[0172] Image feature enhancement module 504 based on gravitational field model: connected to module 502, it integrates a neighborhood potential energy calculation unit to perform physical filtering on the image according to the set radius of action and output a gravitational field enhanced image with high signal-to-noise ratio;

[0173] The weight matrix construction module based on multi-scale Gaussian prior 506: This module works independently of the image data stream and mainly generates standardized physical prior templates based on network configuration parameters to provide initialization parameters for subsequent networks;

[0174] The first-layer structure construction, initialization, and feature output module 508 of the detection network is the data aggregation center of the device. It simultaneously receives the enhanced image from module 504 and the weight matrix from module 506. This module completes the extraction and adaptive fusion of physical features through its internal multi-scale convolutional array and channel attention units, and finally outputs a weighted feature map to the subsequent target detection backbone network.

[0175] For specific limitations regarding the Gaussian kernel matching-based weak target detection preprocessing device, please refer to the limitations of the Gaussian kernel matching-based weak target detection preprocessing method described above, which will not be repeated here. Each module in the aforementioned Gaussian kernel matching-based weak target detection preprocessing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0176] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0177] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A preprocessing method and apparatus for weak target detection based on Gaussian kernel matching, characterized in that, The method includes: Step S100: Radar raw echo, global parameter input, and radar echo feature mapping; acquire two-dimensional time-domain radar raw echo data collected by the radar receiver, and input global configuration parameters; the global configuration parameters include: radar signal processing related parameters, gravitational field enhancement algorithm parameters, and detection network related parameters; wherein, the radar signal processing related parameters include sampling rate, pulse repetition frequency, number of range sampling points, number of pulses, and speed of light; the gravitational field enhancement algorithm parameters include gravitational field action coefficient and gravitational field action radius; the detection network related parameters include the height and width of the network input image, the number of Gaussian template scales, Gaussian variance, convolution kernel size, and preset perturbation parameter matrix; perform two-dimensional discrete Fourier transforms of the radar raw echo data in the range and Doppler directions according to the radar signal processing related parameters to obtain a frequency domain complex matrix, calculate the logarithmic modulus of the frequency domain complex matrix to obtain the amplitude spectrum and resample it to the network input size, and use the Jet color mapping function to convert the amplitude spectrum into a three-channel radar raw feature image; Step S200: Image feature enhancement based on the gravitational field model; acquire the original radar feature image output in step S100, and construct a two-dimensional gravitational field model on the original radar feature image according to the gravitational field action coefficient and gravitational field action radius input in step S100; calculate the neighborhood gravitational potential energy and the internal stress of the pixel itself at each pixel in the image, superimpose the two to obtain the enhanced single-channel intensity value, and use the Jet color mapping function to convert it into a three-channel gravitational field enhanced feature image; Step S300: Construct a weight matrix based on multi-scale Gaussian prior; according to the number of Gaussian template scales, Gaussian variance and convolution kernel size input in step S100, construct a corresponding two-dimensional Gaussian distribution template for each scale; perform coordinate discretization sampling on the template to obtain the original weight matrix, and perform energy normalization processing on the matrix to generate multiple sets of Gaussian initialization weight matrices of different sizes. Step S400: Construct, initialize, and output features of the first layer structure of the detection network; construct a multi-scale Gaussian matching module containing multiple parallel convolutional branches and a subsequent channel weighted fusion module, and use it as the first layer structure of the target detection network; use the multiple sets of Gaussian initialization weight matrices generated in step S300 to assign and initialize the weight parameters of each convolutional branch; receive the gravitational field enhancement feature image output in step S200, input it into the multi-scale Gaussian matching module, extract multi-scale features through parallel convolution operations, and perform weighted fusion by calculating the importance weights of each scale through the channel weighted fusion module, and output the weighted fused feature map as the input data for the subsequent target detection backbone network.

2. The method according to claim 1, characterized in that, Step S100 specifically includes: Step S101: Parameter Input and Parsing; Receive externally input global configuration parameters and parse them to obtain the following variables: radar signal processing related parameters (including sampling rate F). s Pulse repetition frequency f r Number of sampling points N in the distance direction rg Number of pulses N pulse ); Gravitational field enhancement algorithm parameters (including gravitational field action coefficient m, gravitational field radius R); Detection network related parameters (including height H and width W of the network input image, number of Gaussian template scales K, Gaussian variance {σ1, σ2, ..., σ...) K }, kernel size S k Preset perturbation parameter matrix ΔW (k) ); Step S102: Radar echo acquisition; acquire the two-dimensional time-domain raw radar echo signal S(n, m) collected by the externally input radar receiver, where n is the fast time sampling index (0 ≤ n < N). rg ), m is the slow-time pulse index (0≤m<N) pulse ); Step S103: Two-dimensional transformation; Perform two-dimensional discrete Fourier transforms in the range and Doppler directions on the original radar echo S(n, m) obtained in step S102 to obtain the frequency domain complex matrix D(k, l): Where D(k, l) represents the complex value of the frequency domain complex matrix at the k-th distance frequency unit and the l-th Doppler frequency unit; ∑ is the summation symbol, indicating the accumulation operation of values ​​within a specified index range; m is the pulse index variable in the slow time dimension, with a value range of 0 to N. pulse -1:N pulse The number of pulses parsed in step S101 represents the total length of the slow time dimension; n is the sampling point index variable in the fast time dimension, with a value ranging from 0 to N. rg -1; N rg S(n, m) represents the total length of the fast time dimension, which is the number of range sampling points resolved in step S101; S(n, m) is the complex sample value of the radar two-dimensional time-domain echo signal at the nth fast time sampling point and the mth pulse; e is the base of the natural logarithm (Euler's number); j is the imaginary unit, satisfying j 2 =-1; π is the value of a circle. Step S104: Modulus calculation and size resampling; Calculate the logarithmic modulus A(k,l) of the frequency domain complex matrix D(k,l) obtained in step S103: A(k,l)=20log 10 (|D(k,l)|+ò) Where ò is a small constant to prevent logarithmic singularity, and |·| represents the modulo operation of complex numbers; Using the nearest neighbor interpolation algorithm to calculate N rg ×N pulse The logarithmic modulus matrix A of dimension H×W is resampled into a matrix A′ of dimension H×W; Step S105: Normalization; Perform linear normalization on the matrix A′ after resampling in step S104 to obtain V(i,j), the calculation formula is as follows: Where V(i,j) is the normalized pixel value, with a value range of [0,1]; (i,j) is the image coordinate index, satisfying 0≤i<H, 0≤j<W; A′(i,j) is the value of the resampled logarithmic modulus matrix in the i-th row and j-th column; min(A′) is the minimum value of all elements in matrix A′; max(A′) is the maximum value of all elements in matrix A′. Step S106: Pseudo-color mapping; using the Jet color map mapping function, the normalized V(i,j) from step S105 is mapped to red (R), green (G), and blue (B) three-channel pixel values ​​to generate the original radar feature image I. The calculation formula is as follows: I R (i,j)=clamp(1.5-|4·V(i,j)-3|,0,1) I G (i,j)=clamp(1.5-|4·V(i,j)-2|,0,1) I B (i,j)=clamp(1.5-|4·V(i,j)-1|,0,1) Among them, I R I G I B These represent the pixel intensity values ​​of the red, green, and blue channels, respectively; clamp(x, a, b) is the truncation function, defined as: return a when x < a, return b when x > b, otherwise return x; |·| represents the absolute value operation; 1.5, 4, 3, 2, 1 are all fixed piecewise linear function coefficients of the Jet color mapping algorithm.

3. The method according to claim 1, characterized in that, Step S200 specifically includes: Step S201: Determine the gravitational field parameters; obtain the gravitational field interaction coefficient m and gravitational field radius R obtained in step S101; for any pixel (i, j) in the original radar feature image I, determine its neighborhood set Ω with radius R. (i,j) Define h(i,j) as the grayscale intensity value of pixel (i,j). In order to convert the three-channel color information into a single-channel physical scalar field that can characterize the energy intensity of the target, the standard brightness weighted formula is used for calculation: h(i,j)=0.299×I R (i,j)+0.587×I G (i,j)+0.114×I B (i,j) Among them, I R (i, j), I G (i, j), I B (i, j) are the pixel values ​​of the red, green and blue channels of the original radar feature image I(i, j) generated in step S106 at coordinates (i, j); the coefficients 0.299, 0.587 and 0.114 are standard grayscale conversion weights that conform to human visual perception and can preserve the energy distribution characteristics in Jet pseudo-color mapping to the greatest extent. Step S202: Calculate the neighborhood gravitational potential energy; traverse the neighborhood set Ω determined in step S102. (i，j) For all pixels (u, v), calculate the sum G of their gravitational contributions to the center pixel (i, j). ò (i, j), its calculation formula is: Among them, G ò (i, j) represents the sum of the gravitational potential energy in the neighborhood of pixel (i, j); Ω (i,j) Let (i, j) represent the set of neighboring pixels centered at (i, j) with radius R; m is the gravitational field coefficient, used to adjust the intensity of gravitational enhancement; h(i, j) is the original grayscale value of the center pixel (i, j); h(u, v) is the original grayscale value of a pixel (u, v) within the neighborhood calculated in step S201; d is the Euclidean distance between pixel (i, j) and (u, v), calculated using the formula: ∑ is the summation symbol, which indicates that the values ​​within a specified index range are summed. Step S203: Calculate internal stress; calculate internal stress F using the pixel's own strength. In (i, j), its calculation formula is: F In (i,j)=m·[h(i,j)] 2 Among them, F In (i,j) represents the internal stress at pixel (i,j), used to maintain the sharpness of the target edge; m is the gravitational field coefficient; h(i,j) is the original grayscale intensity value of pixel (i,j) calculated in step S201. Step S204: Superimpose to generate enhanced images; calculate gravitational field enhancement feature image I gray The new intensity value of the corresponding pixel is calculated using the following formula: I gray (i,j)=G ò (i,j)+F In (i,j) Among them, I gray (i, j) represents the pixel value of the enhanced image at coordinates (i, j); G ò (i, j) represents the neighborhood gravitational potential energy calculated in step S202; F In (i, j) represents the internal stress calculated in step S203. According to step S105, I is... gray After linear normalization, the Jet color map mapping function described in step S106 is used again to convert it into a three-channel gravitational field enhancement feature image I. ò The image has dimensions of H×W×3 and serves as the input data for the target detection neural network in step S400.

4. The method according to claim 1, characterized in that, Step S300 specifically includes: Step S301: Parameter parsing; Obtain the Gaussian template scale number K and convolution kernel size S from the detection network parameters parsed in step S101. k (S k (where the number is odd) and K groups of Gaussian variance parameters {σ1, σ2, ..., σ... K }; Step S302: Discretize sampling; for the k-th scale (1≤k≤K), traverse the row index r and column index c of the convolution kernel matrix (0≤r, c<S k ), calculate its coordinates (x, y) relative to the center of the convolution kernel: Where x and y are coordinates relative to the center of the convolution kernel; S k The convolution kernel size is obtained in step S301; r and c are the row and column indices of the convolution kernel matrix. This indicates the floor function; Substituting (x, y) into the two-dimensional Gaussian formula yields the unnormalized discrete weight elements. The calculation formula is as follows: in, σ represents the original weight values ​​of the Gaussian template at the k-th scale at matrix coordinates (r, c); exp is an exponential function with the natural constant e as the base; x and y are the horizontal and vertical distances relative to the center of the convolution kernel; σ k The k-th Gaussian variance parameter obtained in step S301 is used to control the distribution width of the Gaussian function; Step S303: Energy normalization; calculate the discrete weight matrix obtained in step S302. Sum of elements k The matrix is ​​then normalized to obtain the initial weight matrix. The calculation formula is as follows: Among them, Sum k It is the sum of all elements of the k-th unnormalized weight matrix; The normalized initial weight values ​​satisfy the condition that the sum of the matrix elements is 1; The discrete weight matrix is ​​calculated in step S302.

5. The method according to claim 1, characterized in that, Step S400 specifically includes: Step S401: Construct a parallel structure; build a multi-scale Gaussian matching module before the backbone network of the object detection network. This module contains K parallel convolutional branches; the kernel size of the k-th branch is configured as S. k And set the zero-fill parameter P k The calculation formula is as follows: Among them, P k This represents the number of layers where the k-th branch pads the edges of the input image with zeros, used to ensure that the feature map space size of all branch outputs remains constant; S k The kernel size obtained in step S301; This indicates the floor function; Step S402: Weight synthesis and assignment; Obtain the preset perturbation parameter matrix parsed in step S101, and superimpose it with the initial weight matrix of the corresponding scale generated in step S303 to calculate the final convolution kernel weights: in, The fixed Gaussian prior initial weights generated in step S300; ΔW (k) The preset perturbation parameter matrix, analyzed in step S101, is used to fit the shape deviation of the real radar target relative to the standard Gaussian model. The final convolution kernel weights are expanded by the number of channels and then assigned to the weight parameters of the corresponding convolution branches, serving as fixed convolution kernel parameters. Step S403: Feature extraction and stitching; the gravitational field enhancement feature image I ò Simultaneously, the input is fed into K parallel convolution branches, and the gravitational field enhancement feature image is convolved using the Gaussian initialization weight matrix assigned in step S402 to obtain K sets of feature maps F1, F2, ..., Fn at different scales. K The K sets of feature maps are concatenated along the channel dimension to form a multi-scale feature tensor F. concat ; Step S404: Adaptive scale weighting; after the concatenated multi-scale feature tensor, a channel weighted fusion module is connected; firstly, the multi-scale feature tensor F formed in step S403 is processed... concat Global average pooling is used to obtain the channel descriptor vector z, and the formula for calculating the c-th element is: Among them, z c Let C be the global statistical feature of the c-th channel, where 1 ≤ c ≤ C. total Scale feature tensor F concat Total number of channels C total The value is K; H and W are the height and width of the feature map; F concat (i, j, c) represents the values ​​of the feature tensor at coordinates (i, j) and channel c; Next, two fully connected layers and a sigmoid activation function are used to generate a weight vector α, which is calculated as follows: α=σ(W2δ(W1z)) in, σ represents the output channel attention weight vector; σ is the Sigmoid function; δ is the ReLU activation function; The above is the channel descriptor vector; Here is the weight matrix of the first dimension-reduced fully connected layer, and r is the dimension reduction ratio; This is the weight matrix for the second fully connected layer with increased dimensionality. Finally, the weight vector α and the feature tensor F are compared. concat Perform channel-by-channel multiplication and output the weighted fused feature map.

6. A preprocessing device for weak target detection based on Gaussian kernel matching, characterized in that, The device includes: The radar raw echo, global configuration parameter input, and radar echo feature mapping module are used to acquire two-dimensional time-domain radar raw echo data collected by the radar receiver and input global configuration parameters. The global configuration parameters include radar signal processing related parameters, gravitational field enhancement algorithm parameters, and detection network related parameters. The module is also used to perform two-dimensional discrete Fourier transforms in the range and Doppler directions on the radar raw echo data according to the radar signal processing related parameters to obtain a frequency domain complex matrix, calculate the logarithmic modulus of the frequency domain complex matrix to obtain the amplitude spectrum, resample it to the network input size, and use the Jet color mapping function to convert the amplitude spectrum into a three-channel radar raw feature image. The image feature enhancement module based on the gravitational field model is used to acquire the original radar feature image and construct a two-dimensional gravitational field model on the original radar feature image according to the gravitational field action coefficient and gravitational field action radius in the gravitational field enhancement algorithm parameters; calculate the neighborhood gravitational potential energy and the internal stress of the pixel itself in each pixel in the image, superimpose the two to obtain the enhanced single-channel intensity value, and use the Jet color mapping function to convert it into a three-channel gravitational field enhanced H×W feature image; The weight matrix construction module based on multi-scale Gaussian prior is used to construct a corresponding two-dimensional Gaussian distribution template for each scale according to the number of Gaussian template scales, Gaussian variance and convolution kernel size in the relevant parameters of the detection network; the template is subjected to coordinate discretization sampling to obtain the original weight matrix, and the matrix is ​​subjected to energy normalization processing to generate multiple sets of Gaussian initial weight matrices of different sizes. The detection network's first-layer structure construction, initialization, and feature output module is used to construct a multi-scale Gaussian matching module containing multiple parallel convolutional branches and a subsequent channel weighted fusion module, which serves as the first-layer structure of the target detection network. The final convolutional kernel weights, generated by superimposing the multiple sets of Gaussian initialization weight matrices with the preset perturbation parameter matrix, are used to initialize the weight parameters of each convolutional branch. The gravitational field enhancement feature image is received and input into the multi-scale Gaussian matching module. Multi-scale features are extracted through parallel convolution operations, and the importance weights of each scale are calculated and weighted fused using the channel weighted fusion module. The weighted fused feature map is then output as input data for the subsequent target detection backbone network.

Images