Sea surface low-speed small target detection method based on three-dimensional feature gridding

Abstract

The invention discloses a sea surface low-speed small target detection method based on three-dimensional feature gridding, and the method comprises the steps: firstly, providing an improved feature extraction method on the basis of three traditional features, and improving the feature difference of the three traditional features in combination with a whitening thought; secondly, a three-dimensional feature gridding detector is designed, an optimal clutter distribution fitting envelope is obtained by uniformly dividing a spatial three-dimensional grid and combining a closed operation, the optimal clutter distribution fitting envelope is used as a judgment threshold of sea clutter and a target, and the accuracy and real-time performance of sea target detection are improved. The method solves the problems that an existing three-feature method is low in feature distinction degree and poor in judgment threshold fitting, a more accurate judgment area can be obtained, and target detection performance and real-time performance are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a method for detecting slow small targets on the sea surface based on three-dimensional feature gridding, which can realize high-reliability and rapid detection of slow small targets under a sea clutter background. Background Art

[0002] As an active microwave target detection device, radar plays an irreplaceable role in the three-dimensional marine surveillance system. However, when detecting low-altitude, slow-moving, small targets, the radar works in a low-grazing angle mode. The small targets take advantage of the natural advantages of sea clutter to make their echoes submerged in strong clutter, greatly increasing the difficulty of radar detection.

[0003] At present, the commonly used sea detection methods can be divided into two categories: energy detection and feature detection. Most energy detection methods use the difference between the target and background clutter energy to achieve the purpose of sea surface target detection. However, for low observable targets, the target echo energy is weak and highly overlaps with the main clutter area in the time domain and frequency domain. It is difficult to achieve robust and robust target detection only by a single energy detection method; the feature detection method maps the echo signal to different representation spaces, extracts robust and separable features in the representation space, and uses feature differences to achieve highly reliable detection of the target. The feature detection method provides an effective solution for sea target detection, and its core lies in the extraction of differentiated features and the design of detectors. In terms of feature extraction, the existing feature extraction mainly includes the traditional three features obtained in the one-dimensional time domain / frequency domain (relative average amplitude RAA, relative Doppler peak height RDPH, relative vector entropy RVE), and the time-frequency three features obtained in the time-frequency two-dimensional domain (ridge accumulation, maximum connected area size, number of connected areas), etc. Among them, the time-frequency three features can achieve better results, but when extracting features, the one-dimensional time domain echo signal needs to be transformed into the time-frequency two-dimensional domain, and the computational complexity limits its application in practice; in terms of feature detectors, the "convex hull" design method is mainly used, but due to the sparse and incomplete characteristics of the target samples in the echo, and the difficulty of the convex hull to fully fit the distribution characteristics of the clutter in the three-dimensional feature space, a large number of targets are missed. Summary of the invention

[0004] In view of the problems of low feature distinguishability and poor decision threshold fitting of the existing three-feature method, the purpose of the present invention is to provide a slow small target detection method on the sea surface based on three-dimensional feature gridding, so as to obtain a more accurate decision area, thereby improving the target detection performance and real-time performance.

[0005] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0006] A method for detecting slow small targets on the sea surface based on three-dimensional feature gridding, comprising:

[0007] For the pulse radar echo signals of known targets, the echo signals are sorted and arranged according to pulses to obtain a two-dimensional echo matrix; in the two-dimensional echo matrix, each row of data represents a sampling point of a pulse, and each column of data represents a sampling point of each range gate; each pulse is pre-processed to obtain the time spectrum data of each range gate data;

[0008] The time spectrum data of each range gate is transformed into Doppler spectrum data through fast Fourier transform, and then each range gate is used as the range gate to be whitened, and the range gates on both sides of it are used as reference range gates, and the median and standard deviation of each Doppler frequency point in the range gate to be whitened are calculated; then, the Doppler frequency point of the range gate to be whitened is subtracted from the corresponding Doppler frequency point median and divided by the standard deviation to obtain the whitened result, and the whitened result is inversely fast Fourier transformed to restore it to time spectrum data;

[0009] The range gates except the one where the target is located are divided into sliding windows. Each window of the range gate division is used as a sea clutter sample unit, and the units on both sides of each sea clutter sample unit are reference units. The three features of RAA feature, RDPH feature and RVE feature are extracted from each sea clutter sample unit.

[0010] After extracting the three features of all sea clutter sample units, the feature vector of each sea clutter sample unit is constructed by the three extracted features, thereby constructing a training sample set; then a three-dimensional feature space is constructed with the three features as coordinate axes, and the feature vector of the sea clutter sample unit is used as the three-dimensional coordinate index to draw the corresponding feature point position of the feature vector of each sea clutter sample unit in the three-dimensional feature space;

[0011] Based on the extreme values ​​of the three features of all sea clutter sample units, a cuboid region is constructed in a three-dimensional coordinate system as a segmentation region, the segmentation region is divided into a three-dimensional grid array, and the number of corresponding feature points of the sea clutter sample units falling into each grid is counted, so as to determine the attributes of each grid; a three-dimensional initial binary image with the same dimension as the three-dimensional grid array is established, and pixel points of the three-dimensional initial binary image are assigned values ​​according to the grid attributes;

[0012] The three-dimensional initial binary image is processed by a morphological closing operation, so that the small holes and disconnected parts in the three-dimensional initial binary image are filled and connected and the outer boundary noise of the image is removed to obtain an updated binary image; the updated binary image is subjected to wild point elimination to obtain a eliminated binary image; and the properties of the grid in the three-dimensional grid array are adjusted according to the eliminated binary image;

[0013] In the actual detection stage, after obtaining the new radar echo signal, its pulse data is pre-processed, and then whitened according to the range gate. All the range gates after whitening are divided into windows, and the range gates in each window are used as units to be detected; the three features of each unit to be detected are extracted, and the feature vector of the unit to be detected is constructed. The feature vector is used as a three-dimensional coordinate index to map to the feature point to be detected in the three-dimensional feature space, and detection is performed according to the position of the feature point to be detected:

[0014] If the feature point to be detected falls outside the space where the three-dimensional grid array is located or falls into a grid of the three-dimensional grid array, but the attribute of the grid is a non-sea clutter grid, then the feature point to be detected is classified as a target; otherwise it is classified as sea clutter.

[0015] Furthermore, the pre-processing includes down-conversion and pulse compression.

[0016] Furthermore, the calculation of the median and standard deviation of each Doppler frequency point in the range gate to be whitened; then, the whitened result is obtained by subtracting the corresponding Doppler frequency point median from the Doppler frequency point of the range gate to be whitened and dividing by the standard deviation, including:

[0017]

[0018] Where x(f) represents the Doppler spectrum data of the range gate to be whitened, x p (f) represents the Doppler spectrum data of the pth reference range gate, P represents the number of reference range gates, median(·) represents the calculated median, med and std represent the median sequence and standard deviation sequence of Doppler frequency point f, respectively; x norm (f) is the range gate Doppler spectrum data after whitening, for x norm (f) Perform inverse fast Fourier transform to restore the time spectrum data x norm (n).

[0019] Furthermore, the extraction formulas of the RAA feature, RDPH feature and RVE feature are as follows:

[0020] The RAA feature extraction formula is:

[0021]

[0022] in, and denote the mean of the energy of the sea clutter sample unit and the reference unit, xnorm(n) and x pnorm (n) is the time spectrum data of the whitened sea clutter sample unit and the corresponding p-th reference unit, n is the time sampling point, N is the total number of time sampling, and P′ represents the number of reference units;

[0023]

[0024] Where, Peak represents the Doppler peak value of each sea clutter sample unit, x norm (f) represents the range gate Doppler spectrum data after whitening, f represents different Doppler frequencies, T is the pulse repetition time, and the corresponding Doppler channel is recorded as f max (x norm (f));

[0025] At this time, DPH can be expressed as:

[0026]

[0027] Wherein, Δ=[-δ1,-δ2]∪[δ2,δ1], δ1 and δ2 represent the maximum Doppler bandwidth occupied by sea clutter and target respectively, and #Δ means the number of Doppler frequency points remaining after excluding the Doppler frequency points that may be occupied by the target in the frequency point area of ​​the sea clutter sample;

[0028] Finally, RDPH features are:

[0029]

[0030] Among them, x pnorm (f) represents the Doppler spectrum reference unit after whitening;

[0031]

[0032] in, Therefore, the RVE characteristics are:

[0033]

[0034] Furthermore, based on the extreme values ​​of the three features of all sea clutter sample units, a rectangular area is constructed in a three-dimensional coordinate system as a segmentation area, the segmentation area is divided into a three-dimensional grid array, and the number of corresponding feature points of the sea clutter sample units falling into each grid is counted, so as to determine the attributes of each grid, including:

[0035] First, the maximum and minimum values ​​of the three features of all sea clutter sample units are calculated respectively and the corresponding endpoint positions are found in the three coordinate axes. The rectangular area surrounded by the six planes corresponding to the six endpoints is used as the segmentation area, which is recorded as P ar ; This area is the smallest rectangular space parallel to the coordinate axis that contains all the corresponding feature points of the sea clutter sample units; the segmentation area P ar The three coordinate dimensions are evenly divided, so that the segmented area P arDivide into a three-dimensional grid array, wherein each grid is set with an initial value num=0 to record the number of feature points corresponding to the sea clutter sample unit falling into the grid;

[0036] Then, the corresponding feature points of the sea clutter sample units are traversed one by one, and the num count of the grid is increased by 1 every time a feature point falls into the grid.

[0037] Furthermore, a three-dimensional initial binary image having the same dimension as the three-dimensional grid array is established, and pixel points of the three-dimensional initial binary image are assigned values ​​according to the grid attributes, including:

[0038] A three-dimensional initial binary image B with the same dimension size is constructed according to the three-dimensional grid array; the grids whose grid num count is greater than 0 are called sea clutter grids, and their corresponding pixels in the three-dimensional initial binary image B are recorded as 1, and the rest are non-sea clutter grids, and the corresponding pixels are recorded as 0.

[0039] Furthermore, the closing operation processing includes:

[0040] The specific process of this operation is to set the structural element "se" to first "expand" the original image B, and then perform "erosion" operation; wherein the structural element "se" is the basic unit of "expand" and "erosion" operations, and the structural element "se" is made to slide in the initial binary image B, and the "se" origin position is set to 1 and 0 according to the definition of "expand" and "erosion"; here, the structural element "se" can also select a rectangular parallelepiped style according to the rectangular parallelepiped structure of the initial binary image B, and the center point of the structural element "se" is used as the origin position; the updated binary image is recorded as B C , B C The grid corresponding to the point with a median value of 1 in the three-dimensional grid array is recorded as a sea clutter grid, and the grid corresponding to the point with a median value of 0 is set to zero and recorded as a non-sea clutter grid.

[0041] Furthermore, the process of removing outliers is as follows:

[0042] For the updated binary image B C The connected areas with a value of 1 are counted, and the connected area with the largest number of grids in the corresponding three-dimensional grid array is called the main connected area. The corresponding grids of the remaining connected areas outside the main connected area in the three-dimensional grid array are set to zero and discarded as wild points; the binary image after eliminating wild points is recorded as J, and the grids corresponding to the points with a value of 1 in J in the three-dimensional grid array are set as sea clutter grids, and the rest are non-sea clutter grids.

[0043] Furthermore, the number of feature points corresponding to the sea clutter sample units in the three-dimensional grid array that are set to zero in the morphological closing operation and the outlier removal operation is counted, and the false alarm rate of the current classifier can be obtained by the number of feature points and the total number of sea clutter sample units.

[0044] A sea surface slow small target detection device comprises a processor, a memory and a computer program stored in the memory; when the processor executes the computer program, the sea surface slow small target detection method based on three-dimensional feature gridding is implemented.

[0045] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the method for detecting slow small targets on the sea surface based on three-dimensional feature gridding is implemented.

[0046] Compared with the prior art, the present invention has the following technical features:

[0047] 1. The traditional three-feature extraction method is improved, and the distinguishability of features in three-dimensional space is improved; 2. Through the three-dimensional grid processing of features, the threshold of the decision device is more closely matched with the clutter distribution, which improves the detection performance under complex clutter background. The algorithm is verified using the public IPIX measured sea clutter data set. Compared with the "traditional three-feature + convex hull" method, the proposed method improves the average detection rate of the target by 20.34% under the same false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is the sea clutter distribution and convex hull decision area;

[0049] Figure 2 It is a schematic diagram of the process of the present invention;

[0050] Figure 3 is a schematic diagram of an original echo matrix in one embodiment of the present invention;

[0051] Figure 4 A schematic diagram of a segmented area in one embodiment of the present invention;

[0052] Figure 5 is a sea clutter grid corresponding to an initial binary image in one embodiment of the present invention;

[0053] Figure 6 is a sea clutter grid corresponding to a binary image after a closing operation in one embodiment of the present invention;

[0054] Figure 7 A sea clutter grid corresponding to a binary image after outliers are removed in one embodiment of the present invention;

[0055] Figure 8 The detection rate comparison between the present invention and the traditional fast convex hull method under the same false alarm rate;

[0056] Fig. 9 The training time comparison between the present invention and the traditional fast convex hull method under the same false alarm rate;

[0057] Fig.10 The detection time comparison between the present invention and the traditional fast convex hull method at the same false alarm rate. DETAILED DESCRIPTION

[0058] In order to further improve the target detection performance in complex and variable environments, the present invention proposes a three-dimensional feature gridded sea surface slow small target detection method. This method first uses the frequency domain whitening method to optimize the extraction of the traditional three features (relative average amplitude, relative Doppler peak height, relative vector entropy), and improves the difference between sea clutter and target features in three-dimensional space; then, using the idea of ​​extreme approximation, a three-dimensional feature gridded detector is designed, and the contour of the clutter distribution is fitted through the grid to overcome the "fitting blank area" (such as the convex hull) caused by the "convex hull" method. Figure 1 ), improving the detection performance in complex scenes. Figure 1 It can be seen that there are large blank areas in the convex hull judgment area that do not match the sea clutter distribution contour. This part will affect subsequent target detection and increase the missed alarm rate.

[0059] See attached Figure 2 The present invention provides a method for detecting slow small targets on the sea surface based on three-dimensional feature gridding, comprising the following steps:

[0060] Step 1: For the pulse radar echo signal of the known target, the echo signal is sorted and arranged according to the pulses to obtain a two-dimensional echo matrix; in the two-dimensional echo matrix, each row of data represents the sampling point of a pulse, and each column of data represents the sampling point of each range gate; each pulse (i.e., each row of data) is pre-processed by down-conversion, pulse compression, etc. to obtain the time spectrum data of each range gate data (i.e., each column of data).

[0061] Feature extraction is a pre-step in detector design. The corresponding detector can only be trained by extracting sea clutter features from sea clutter radar echoes of known targets to obtain a training sample set. In this step, a two-dimensional echo matrix is ​​established for the echo signal received by the pulse radar, and each pulse data is pre-processed for feature extraction; the range gate where the target is located is known; the two-dimensional echo matrix constructed by an embodiment of the present invention is as follows: Figure 3 shown.

[0062] Step 2: First, the time spectrum data of each range gate is converted into Doppler spectrum data through fast Fourier transform, and then each range gate is used as the range gate to be whitened, and the range gates on both sides thereof are used as reference range gates, and the median and improved standard deviation of each Doppler frequency point in the range gate to be whitened are calculated; then, the Doppler frequency point of the range gate to be whitened is subtracted from the corresponding Doppler frequency point median and divided by the improved standard deviation to obtain the whitened result, and the whitened result is inversely fast Fourier transformed to restore it to the time spectrum data.

[0063] The present invention adopts the idea of ​​whitening processing. Before extracting the three characteristics of sea clutter, whitening processing is performed first to enhance the difference between the target and sea clutter characteristics.

[0064] The whitening operation is processed in units of range gates. The range gate data after pre-processing is time spectrum data, expressed as x(n), where n is the time sampling point; x(n) is subjected to fast Fourier transform to obtain Doppler spectrum data, expressed as x(f), where f is the Doppler frequency point; the Doppler spectrum of sea clutter has the same distribution between range gates, so the range gates on both sides of the range gate to be whitened are used as references to calculate the mean and standard deviation of each Doppler frequency point; then, the whitened result is obtained by subtracting the mean of the corresponding frequency point from the Doppler frequency point of the range gate to be whitened and dividing it by the standard deviation, and an inverse fast Fourier transform is performed on it to restore the time spectrum data; in order to avoid errors caused by individual maximum or minimum values, the median can be further used to replace the mean in the mean and standard deviation calculation. The calculation formula for whitening is as follows:

[0065]

[0066] Where x(f) represents the Doppler spectrum data of the range gate to be whitened, x p (f) represents the Doppler spectrum data of the pth reference range gate, where P represents the number of reference range gates, which is set according to actual needs; median(·) represents the median calculation, that is, the sequence consisting of the median values ​​of each Doppler frequency point f in the P reference range gates; med and std represent the median sequence and improved standard deviation sequence of the Doppler frequency point f, respectively; x norm (f) is the range gate Doppler spectrum data after whitening, for x norm (f) Perform inverse fast Fourier transform to restore the time spectrum data x norm (n).

[0067] After all range gates are whitened in the frequency domain according to the above operation, feature extraction can be performed.

[0068] Step 3: Perform sliding window division on the range gates except the range gate where the target is located. Each window of the range gate division is used as a sea clutter sample unit, and the units on both sides of each sea clutter sample unit are reference units. Three features of RAA feature, RDPH feature and RVE feature are extracted from each sea clutter sample unit.

[0069] This scheme performs sliding window processing on all range gates before feature extraction; the sliding window processing is only for the remaining range gates except the range gate where the target is located, that is, the sea clutter range gate, so as to obtain the sea clutter sample unit; the number of windows divided by each range gate is set according to actual needs, which depends on the actual number of pulses contained in each range gate; finally, three features of each sample unit are extracted.

[0070] (1) Extract the relative average amplitude (RAA) feature.

[0071] The RAA feature is extracted from the time domain, which mainly reflects the difference in energy between the target and the sea clutter. From the analysis of the time domain graph, it is known that the range gate containing the target echo has stronger energy, so the RAA calculation formula is as follows:

[0072]

[0073] in, and denote the mean of the energy of the sea clutter sample unit and the reference unit, xnorm(n) and x pnorm (n) is the time spectrum data of the whitened sea clutter sample unit and the corresponding p-th reference unit, n is the time sampling point, N is the total number of time sampling, and P′ represents the number of reference units.

[0074] (2) Extract relative Doppler peak height (RDPH) features.

[0075] The RDPH feature is extracted from the Doppler frequency domain, which mainly reflects the difference in energy between the target and the sea clutter Doppler amplitude spectrum (DAS). Through the analysis of the frequency domain diagram, it can be seen that the target energy on the DAS is generally higher and has a sharp peak, while the sea clutter peak is generally lower than the target and presents a gentle peak. Therefore, the Doppler peak height (Doppler Peak Height, DPH) in each sea clutter sample unit is first calculated, and the calculation formula is as follows:

[0076]

[0077] Where, Peak represents the Doppler peak value of each sea clutter sample unit, x norm (f) represents the whitened range gate Doppler spectrum data, f represents different Doppler frequencies, T is the pulse repetition time (PRT), and the corresponding Doppler channel is recorded as f max (x norm (f)).

[0078] At this time, DPH can be expressed as:

[0079]

[0080] Wherein, Δ=[-δ1,-δ2]∪[δ2,δ1], δ1 and δ2 represent the maximum Doppler bandwidth occupied by sea clutter and target respectively, and #Δ means the number of Doppler frequency points remaining after excluding the Doppler frequency points that may be occupied by the target in the frequency point area of ​​the sea clutter sample.

[0081] Finally, RDPH is defined as:

[0082]

[0083] Among them, x pnorm (f) represents the Doppler spectrum reference unit after whitening.

[0084] (3) Extract relative vector entropy (RVE) features.

[0085] The RVE feature is extracted from the Doppler frequency domain, which mainly reflects the concentration of target and sea clutter energy. Through the analysis of the frequency domain diagram, it can be seen that the target energy on DAS is more concentrated than the sea clutter energy; the vector entropy (VE) of the data can be expressed as:

[0086]

[0087] in, Therefore, RVE can be expressed as:

[0088]

[0089] Step 4: After extracting the three features of all sea clutter sample units, construct the feature vector F = [RAA, RDPH, RVE] of each sea clutter sample unit through the extracted three features, so as to construct a training sample set; then construct a three-dimensional feature space with the three features as coordinate axes, and use the feature vector of the sea clutter sample unit as the three-dimensional coordinate index to draw the corresponding feature point position of the feature vector of each sea clutter sample unit in the three-dimensional feature space.

[0090] Step 5: Based on the extreme values ​​of the three features of all sea clutter sample units, a rectangular parallelepiped area is constructed in the three-dimensional coordinate system as a segmentation area, and the segmentation area is divided into a three-dimensional grid array, and the number of corresponding feature points of the sea clutter sample units falling into each grid is counted, so as to determine the attributes of each grid (sea clutter grid or non-sea clutter grid); a three-dimensional initial binary image with the same dimension as the three-dimensional grid array is established, and pixel points of the three-dimensional initial binary image are assigned values ​​according to the grid attributes.

[0091] After obtaining the three-dimensional feature space distribution of the feature points corresponding to the sea clutter sample unit, the detector design is carried out based on this; the three-feature gridding process is first performed, and the processing flow is as follows:

[0092] First, the maximum and minimum values ​​of the three features of all sea clutter sample units are calculated respectively and the corresponding endpoint positions are found in the three coordinate axes. The rectangular area surrounded by the six planes corresponding to the six endpoints is used as the segmentation area, which is recorded as P ar ,like Figure 4 As shown; this area is the smallest rectangular space parallel to the coordinate axis containing all the corresponding feature points of the sea clutter sample units; the segmentation area P ar The three coordinate dimensions are evenly divided (for example, the RAA feature corresponding dimension is evenly divided into intervals determined by its maximum and minimum values, and the number of divisions can be selected as 200 based on experience), so that the segmented area P ar The grid is divided into a three-dimensional grid array, wherein each grid is set with an initial value num=0 to record the number of feature points corresponding to the sea clutter sample unit falling into the grid.

[0093] Then, the corresponding feature points of the sea clutter sample units are traversed one by one, and the num count of the grid is increased by 1 every time a feature point falls into the grid.

[0094] Finally, a three-dimensional initial binary image B with the same dimension size is constructed based on the three-dimensional grid array; the grids with grid num count greater than 0 are called sea clutter grids, and their corresponding pixels in the three-dimensional initial binary image B are recorded as 1, and the rest are non-sea clutter grids, and the corresponding pixels are recorded as 0; the decision formula is as follows:

[0095]

[0096] Wherein, x, y, and z represent the index values ​​of the three dimensions of the three-dimensional initial binary image B, and their value ranges are x=1,2,…X, y=1,2,…Y, and z=1,2,…Z, respectively. X, Y, and Z are the grid numbers of the three dimensions. The initial binary image B is as follows: Figure 5 As shown in the figure, the marked grids are sea clutter grids, and there are characteristic points corresponding to sea clutter sample units inside them, while non-sea clutter grids are not drawn. Figure 5It can be seen that the grids are not continuous at this time, and there are more gaps and discrete grids.

[0097] Step 6, performing morphological closing operations on the three-dimensional initial binary image, so that the small holes and disconnected parts in the three-dimensional initial binary image are filled and connected and the outer boundary noise of the image is removed to obtain an updated binary image; performing wild point elimination on the updated binary image to obtain a eliminated binary image; and adjusting the properties of the grid in the three-dimensional grid array according to the eliminated binary image.

[0098] (1) Closing operation processing in morphology.

[0099] The distribution of the feature points corresponding to the sea clutter sample units in the three-dimensional space is non-uniform. At the edge with low probability density, the distance between the sea clutter features is large, resulting in more gaps and discrete points in the initial binary image, which has a serious impact on the judgment. To solve this problem, the closing operation in mathematical morphology is used here to process the initial binary image B.

[0100] The specific process of this operation is to set the structure element "se" to first "dilate" the original image B, and then perform "erosion" operations; wherein, the structure element "se" is the basic unit of the "dilation" and "erosion" operations, and the structure element "se" is allowed to slide in the initial binary image B, and the "dilation" and "erosion" definitions are used to set 1 and 0 operations at the origin position of "se"; here, the structure element "se" can also select a rectangular style according to the rectangular structure of the initial binary image B, and the center point of the structure element "se" is used as the origin position.

[0101] The "dilation" operation can fill and connect small holes and disconnected parts in the binary image; the "erosion" operation can remove the noise at the outer boundary of the image. The closing operation can fill the gaps and connect the breakpoints while making the boundary smoother and more continuous, more in line with the original sea clutter distribution contour; the updated binary image is denoted as B C , B C The grid corresponding to the point with a median value of 1 in the three-dimensional grid array is recorded as the sea clutter grid, and the grid corresponding to the point with a median value of 0 is set to zero and recorded as the non-sea clutter grid, such as Figure 6 As shown in the figure, most of the sea clutter grids have been connected into a whole, and only a few sea clutter grids are still far away from the main body. At the same time, the outline of the main body of the sea clutter grid is closer to the sea clutter distribution.

[0102] (2) Eliminate wild points.

[0103] Binary image B after morphological closing operation C There may still be a small number of points with a value of 1 that are far away from the main body. The grids corresponding to these points in the three-dimensional grid array are called discrete grids, such as Figure 6 As shown in Figure 2. Since this type of grid is far away from the main clutter distribution, it can be discarded as a wild point; the specific operation is to update the binary image B C The connected areas with a value of 1 are counted, and the connected area with the largest number of grids in the corresponding three-dimensional grid array is called the main connected area. The corresponding grids of the remaining connected areas outside the main connected area in the three-dimensional grid array are set to zero and discarded as wild points; the binary image after removing the wild points is recorded as J, and the grids corresponding to the points with a value of 1 in J in the three-dimensional grid array are set as sea clutter grids, and the rest are non-sea clutter grids, such as Figure 7 As shown in the figure, this area is the main connected area of ​​the binary image after the closing operation, and the connection mode is 18-connected. The sea clutter grids that are discrete outside the main connected area are set to zero as wild points, thereby controlling the false alarm rate.

[0104] Step 7: In the actual detection stage, after obtaining the new radar echo signal, the pulse data is pre-processed, and then whitened according to the range gate, and all the range gates after the whitening process are divided into windows, and the range gates in each window are used as the units to be detected;

[0105] Extract the three features of each unit to be detected, construct the feature vector of the unit to be detected, map the feature vector as the three-dimensional coordinate index to the feature point to be detected in the three-dimensional feature space, and perform detection according to the position of the feature point to be detected:

[0106] If the feature points to be detected fall in the space P of the three-dimensional grid array ar If it is outside or falls into a grid in the three-dimensional grid array, but the attribute of the grid is a non-sea clutter grid, then the feature point to be detected is classified as a target; otherwise it is classified as sea clutter; the judgment formula is as follows:

[0107]

[0108] Wherein, CUT represents the feature point to be detected.

[0109] The false alarm rate of the classifier is set based on the number of feature points corresponding to the grid sea clutter sample units that are set to zero in the three-dimensional grid array during the closing operation processing and the outlier elimination process.

[0110] The number of feature points corresponding to the sea clutter sample units in the three-dimensional grid array that is set to zero in the above morphological closing operation and the outlier removal operation is counted, and the false alarm rate of the current classifier can be obtained by dividing the number of feature points by the total number of sea clutter sample units.

[0111] The number of discarded grid points is related to the size of the structural element "se", so the false alarm rate of the detector can be controlled by controlling the volume of the structural element "se"; the larger the volume of the structural element "se", the lower the false alarm rate; the smaller the volume, the greater the false alarm rate.

[0112] Example:

[0113] This embodiment uses IPIX publicly measured sea clutter data, which provides echo data of 14 range gates. Each range gate collects 131072 pulse echo signals and two polarization channels H and V. In order to expand the training and test data sets, a sliding window method is used to construct three-feature data, in which the window length is 1024 pulses. Each windowing can extract the features of 14 range gates, and the sliding window step length is 128 pulses.

[0114] In this embodiment, the maximum and minimum values ​​of the three features of the sea clutter training sample set are calculated respectively, and are denoted as RAA max , RAA min ,RDPH max ,RDPH min , RVE max , RVE min . Use each group of maximum and minimum values ​​as endpoints to frame the segmentation area P ar ,like Figure 4 Each dimension is evenly divided into 200 parts to obtain a three-dimensional grid array of size 200×200×200. Each grid is set with an initial value num=0 to record the number of sea clutter samples falling into the grid.

[0115] Set the structure element se and define it as a three-dimensional array of size 60×60×60 (the size is chosen to be 30% of the length of each dimension based on experience).

[0116] The method proposed in the present invention is compared with the fast convex hull algorithm, and the detection rate, detector training time and detection time under the same false alarm rate are compared. The comparison results are as follows: Figure 8-10 As shown. Figure 8 It can be seen that in the 10 groups of IPIX data, under the same false alarm rate, the detection rate of the method of the present invention is higher than that of the fast convex hull method, and the average performance improvement is 20.34%. Fig. 9 It can be seen that under the same false alarm rate, the average time consumption of the detector training of the method of the present invention is faster than the average time consumption of the fast convex hull algorithm. The training time consumption of the fast convex hull algorithm increases exponentially with the increase of the set false alarm rate, so the training time consumption of 10 sets of data will vary greatly. The training time consumption of the detector of the present invention is mainly reflected in the statistics of landing points, closed operations and connected area calculations. These operations and calculations are relatively simple, and the time consumption is mainly related to the number of sea clutter samples, so the overall difference is not large. Fig.10It can be found that the detection time of this method is significantly faster than that of the fast convex hull algorithm, with a difference of three orders of magnitude.

[0117] In summary, whitening can increase the distinction between sea clutter and target features, making the gap between the two in the three-dimensional feature space larger; with the help of refined fitting of the three-dimensional grid, the judgment area is more closely aligned with the sea clutter distribution contour, improving the detection performance. Through experiments with measured data, it is found that the method proposed in the present invention is superior to the traditional fast convex hull algorithm in detection rate, detector training time and detection time, and has greater practical value.

[0118] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for detecting slow small targets on the sea surface based on three-dimensional feature gridding, characterized in that: include: For pulse radar echo signals of known targets, the echo signals are sorted and arranged according to pulses to obtain a two-dimensional echo matrix; In the two-dimensional echo matrix, each row of data represents the sampling point of a pulse, and each column of data represents the sampling point of each range gate; Perform pre-processing on each pulse to obtain the time spectrum data of each range gate data; The time spectrum data of each range gate is transformed into Doppler spectrum data through fast Fourier transform, and then each range gate is used as the range gate to be whitened, and the range gates on both sides of it are used as reference range gates, and the median and standard deviation of each Doppler frequency point in the range gate to be whitened are calculated; then, the Doppler frequency point of the range gate to be whitened is subtracted from the corresponding Doppler frequency point median and divided by the standard deviation to obtain the whitened result, and the whitened result is inversely fast Fourier transformed to restore it to time spectrum data; The range gates except the one where the target is located are divided into sliding windows. Each window of the range gate division is used as a sea clutter sample unit, and the units on both sides of each sea clutter sample unit are reference units. The three features of RAA feature, RDPH feature and RVE feature are extracted from each sea clutter sample unit. After extracting the three features of all sea clutter sample units, the feature vector of each sea clutter sample unit is constructed by the three extracted features, thereby constructing a training sample set; then a three-dimensional feature space is constructed with the three features as coordinate axes, and the feature vector of the sea clutter sample unit is used as the three-dimensional coordinate index to draw the corresponding feature point position of the feature vector of each sea clutter sample unit in the three-dimensional feature space; Based on the extreme values ​​of the three features of all sea clutter sample units, a cuboid region is constructed in a three-dimensional coordinate system as a segmentation region, the segmentation region is divided into a three-dimensional grid array, and the number of corresponding feature points of the sea clutter sample units falling into each grid is counted, so as to determine the attributes of each grid; a three-dimensional initial binary image with the same dimension as the three-dimensional grid array is established, and pixel points of the three-dimensional initial binary image are assigned values ​​according to the grid attributes; The three-dimensional initial binary image is processed by a morphological closing operation, so that the small holes and disconnected parts in the three-dimensional initial binary image are filled and connected and the outer boundary noise of the image is removed to obtain an updated binary image; the updated binary image is subjected to wild point elimination to obtain a eliminated binary image; and the properties of the grid in the three-dimensional grid array are adjusted according to the eliminated binary image; In the actual detection stage, after obtaining the new radar echo signal, its pulse data is pre-processed, and then whitened according to the range gate. All the range gates after whitening are divided into windows, and the range gates in each window are used as units to be detected; the three features of each unit to be detected are extracted, and the feature vector of the unit to be detected is constructed. The feature vector is used as a three-dimensional coordinate index to map to the feature point to be detected in the three-dimensional feature space, and detection is performed according to the position of the feature point to be detected: If the feature point to be detected falls outside the space where the three-dimensional grid array is located or falls into a grid of the three-dimensional grid array, but the attribute of the grid is a non-sea clutter grid, then the feature point to be detected is classified as a target; otherwise it is classified as sea clutter.

2. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: The calculation of the median and standard deviation of each Doppler frequency point in the range gate to be whitened; Then, the whitened result is obtained by subtracting the corresponding Doppler frequency median from the Doppler frequency of the range gate to be whitened and dividing it by the standard deviation, including: Where x(f) represents the Doppler spectrum data of the range gate to be whitened, x p (f) represents the Doppler spectrum data of the pth reference range gate, P represents the number of reference range gates, median(·) represents the calculated median, med and std represent the median sequence and standard deviation sequence of Doppler frequency point f, respectively; x norm (f) is the range gate Doppler spectrum data after whitening, for x norm (f) Perform inverse fast Fourier transform to restore the time spectrum data x norm (n).

3. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: The extraction formulas of the RAA feature, RDPH feature and RVE feature are as follows: The RAA feature extraction formula is: in, and denote the mean of the energy of the sea clutter sample unit and the reference unit, xnorm(n) and x pnorm (n) is the time spectrum data of the whitened sea clutter sample unit and the corresponding p-th reference unit, n is the time sampling point, N is the total number of time sampling, and P′ represents the number of reference units; Where, Peak represents the Doppler peak value of each sea clutter sample unit, x norm (f) represents the range gate Doppler spectrum data after whitening, f represents different Doppler frequencies, T is the pulse repetition time, and the corresponding Doppler channel is recorded as f max (x norm (f)); At this time, DPH can be expressed as: Wherein, Δ=[-δ1,-δ2]∪[δ2,δ1], δ1 and δ2 represent the maximum Doppler bandwidth occupied by sea clutter and target respectively, and #Δ means the number of Doppler frequency points remaining after excluding the Doppler frequency points that may be occupied by the target in the frequency point area of ​​the sea clutter sample; Finally, RDPH features are: Among them, x pnorm (f) represents the Doppler spectrum reference unit after whitening; in, Therefore, the RVE characteristics are:

4. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: Based on the extreme values ​​of the three features of all sea clutter sample units, a rectangular area is constructed in a three-dimensional coordinate system as a segmentation area, the segmentation area is divided into a three-dimensional grid array, and the number of corresponding feature points of the sea clutter sample units falling into each grid is counted, so as to determine the attributes of each grid, including: First, the maximum and minimum values ​​of the three features of all sea clutter sample units are calculated respectively and the corresponding endpoint positions are found in the three coordinate axes. The rectangular area surrounded by the six planes corresponding to the six endpoints is used as the segmentation area, which is recorded as P ar ; This area is the smallest rectangular space parallel to the coordinate axis that contains all the corresponding feature points of the sea clutter sample units; the segmentation area P ar The three coordinate dimensions are evenly divided, so that the segmented area P ar Divide into a three-dimensional grid array, wherein each grid is set with an initial value num=0 to record the number of feature points corresponding to the sea clutter sample unit falling into the grid; Then, the corresponding feature points of the sea clutter sample units are traversed one by one, and the num count of the grid is increased by 1 every time a feature point falls into the grid.

5. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: Establishing a three-dimensional initial binary image with the same dimension as the three-dimensional grid array, and assigning values ​​to pixel points of the three-dimensional initial binary image according to the grid attributes, including: A three-dimensional initial binary image B with the same dimension size is constructed according to the three-dimensional grid array; the grids whose grid num count is greater than 0 are called sea clutter grids, and their corresponding pixels in the three-dimensional initial binary image B are recorded as 1, and the rest are non-sea clutter grids, and the corresponding pixels are recorded as 0.

6. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: The closing operation processing includes: The specific process of this operation is to set the structural element "se" to first "expand" the original image B, and then perform "erosion" operation; wherein, the structural element "se" is the basic unit of "expand" and "erosion" operations, and the structural element "se" is made to slide in the initial binary image B, and the "se" origin position is set to 1 and 0 according to the definition of "expand" and "erosion"; here, the structural element "se" can also select a rectangular parallelepiped style according to the rectangular parallelepiped structure of the initial binary image B, and the center point of the structural element "se" is used as the origin position; the updated binary image is recorded as B C , B C The grid corresponding to the point with a median value of 1 in the three-dimensional grid array is recorded as a sea clutter grid, and the grid corresponding to the point with a median value of 0 is set to zero and recorded as a non-sea clutter grid.

7. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: The process of outlier removal is as follows: For the updated binary image B C The connected areas with a value of 1 are counted, and the connected area with the largest number of grids in the corresponding three-dimensional grid array is called the main connected area. The corresponding grids of the remaining connected areas outside the main connected area in the three-dimensional grid array are set to zero and discarded as wild points; The binary image after removing wild points is recorded as J. The corresponding grids in the three-dimensional grid array of the points with a value of 1 in J are set as sea clutter grids, and the rest are non-sea clutter grids.

8. The method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to claim 1 is characterized in that: In the statistical morphological closing operation and the outlier removal operation, the number of feature points corresponding to the sea clutter sample units in the three-dimensional grid array is set to zero, and the false alarm rate of the current classifier can be obtained by combining the number of feature points with the total number of sea clutter sample units.

9. A sea surface slow small target detection device, comprising a processor, a memory and a computer program stored in the memory; characterized in that: When the processor executes the computer program, it implements the method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to any one of claims 1 to 8.

10. A computer-readable storage medium, wherein a computer program is stored in the medium; characterized in that: When the computer program is executed by a processor, the method for detecting slow small targets on the sea surface based on three-dimensional feature gridding according to any one of claims 1 to 8 is implemented.

Images