A Comprehensive Feature Extraction Method for RCS Sequences

Abstract

The present invention belongs to the technical field of radar target recognition, and particularly relates to a method for extracting comprehensive features of an RCS sequence. The method of the present invention first performs time-frequency analysis on the training RCS sequence to obtain the corresponding time-frequency spectrogram. Then, the time-frequency spectrogram is divided into non-overlapping sub-blocks, the distribution probability of the binary differential coding sequence of the sub-blocks and the state change probability of adjacent sub-blocks are statistically analyzed, and the two are synthesized to form the classification features of the target. Since the distribution probability of the sub-block coding sequence describes the local structure features, and at the same time the state change probability of adjacent sub-blocks characterizes the correlation features between adjacent local structures, the recognition performance of the target is improved. Simulation experiments are carried out on the RCS data of four types of simulated targets, and the experimental results verify the effectiveness of the method.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and particularly relates to a method for extracting comprehensive features of an RCS sequence. Background Art

[0002] In radar target recognition, the subspace method can analyze data from a global perspective, extract the global and local structural features of target data distribution, and achieve good recognition results. However, research shows that in different target data, in addition to the overall global differences, there are still certain local differences in details. If the local structural features representing target details can be extracted, the recognition rate can be improved. Therefore, there is room for further improvement in the recognition performance of the traditional subspace method that can only extract global structural features. Summary of the Invention

[0003] The purpose of the present invention is to propose a comprehensive feature extraction method. By performing sub-block analysis on the time-frequency spectrogram of the training data sequence, on the one hand, the local distribution features of the sub-block structure are extracted, and on the other hand, the change correlation features between adjacent sub-blocks are extracted from the global aspect, and the two types of features are combined to form the classification features of the target, thereby improving the recognition performance of the target.

[0004] The technical solution of the present invention is as follows:

[0005] A method for extracting comprehensive features of an RCS sequence, comprising the following steps:

[0006] S1. Define the j-th training RCS data sequence frame of the i-th type of true or false target as an n-dimensional column vector x ij , 1 ≤ i ≤ C, 1 ≤ j ≤ N i , where N i is the number of frames of the training RCS sequence of the i-th type of true or false target, then the total number of frames of the training RCS sequence

[0007] S2. Obtain the time-frequency spectrogram by short-time Fourier transform: Perform short-time Fourier transform on the training sample data x ij of the i-th type of true or false target to obtain the time-frequency spectrogram S ij :

[0008] S ij = [s ij,km K×M

[0009] where s ij,km represents the element in the time-frequency spectrogram S ij , k = 1, 2,... K; m = 1, 2,... M, k is the row subscript representing the frequency change direction, m is the column subscript representing the time change direction, K is the number of rows of S ij , and M is the number of columns of Sij The number of columns;

[0010] S3. When dividing the sub - blocks, for the spectrogram S ij , obtain the sub - block differential binary coding sequence:

[0011] With a 4x4 window, perform non - overlapping sliding on the time - frequency spectrogram S ij to obtain multiple 4x4 sub - blocks, forming a sub - block set Q ij :

[0012] Q ij = [P ij,1 P ij,2 …P ij,L

[0013] P ij,l = [p ij,l,te 4×4 , t, e = 1, 2, 3, 4

[0014] l = 1, 2, … L

[0015] Among them, P ij,l represents the l - th sub - block matrix in the sub - block set Q ij , p ij,l,te represents the element in the sub - block matrix P ij,l , and L represents the total number of sub - blocks;

[0016] Calculate the mean value of all elements in the sub - block matrix P ij,l

[0017]

[0018] Perform the following encoding on p ij,l in the sub - block P ij,l,te :

[0019]

[0020] Among them, c ij,l,te is the encoding corresponding to p ij,l,te . The encodings of each element in the sub - block form a 0 and 1 binary differential coding sequence c ij,l :

[0021] c ij,l = [c ij,l,11 c ij,l,12 … c ij,l,44

[0022] S4. Extract the local structure features of the sub - block differential binary coding distribution:

[0023] For the differential coding sequence c ij,lAs a binary number, count the number of repeated occurrences of the differential coding sequence values in each sub-block of the entire time-frequency spectrum diagram and form a vector:

[0024] [π ij,1 π ij,2 … π ij,L1

[0025] Among them, sort the differential coding sequence values of the sub-blocks in ascending order. The number of these values is L1, and π ij,1 is the number of repeated occurrences of the first value, and π ij,2 is the number of repeated occurrences of the second value, is the number of repeated occurrences of the L1-th value;

[0026] Obtain the local structure feature vector h of the differential binary coding distribution corresponding to the time-frequency spectrum diagram S ij : ij

[0027]

[0028] S5. Extract the global local correlation features of the state changes between adjacent sub-blocks:

[0029] Evenly divide the distribution interval of the binary coding sequence values of the sub-blocks into 16 sub-intervals. The central values of each sub-interval are v1, v2,... v 16 , and count the number of repeated state changes of the binary coding sequence values of adjacent sub-blocks in the time-frequency spectrum diagram:

[0030]

[0031] Among them, represents the number of repeated occurrences of the coding value of the adjacent sub-block being v f (f = 1, 2... 16) under the condition that the coding value of the current sub-block is v r (r = 1, 2... 16). After normalizing the statistical matrix row by row, obtain the state change feature:

[0032]

[0033] Among them, W ij represents the state change feature matrix;

[0034] S6. Combine the local structure feature h of the binary coding distribution of the sub-blocks ij and the global correlation feature W of the state changes between the sub-blocks ij to obtain the combined feature vector:

[0035] [h ij w ij,1 w ij,2 …w​​ij,16

[0036] Among them, w ij,1 、 w ij,2 、 w ij,16 respectively represent the first row vector, the second row vector, and the sixteenth row vector in the matrix W ij .

[0037] The beneficial effects of the present invention are as follows: First, the present invention performs time-frequency analysis on the training RCS sequence to obtain the corresponding time-frequency spectrogram of the sequence. Then, the time-frequency spectrogram is divided into non-overlapping sub-blocks, and the distribution probability of the binary differential coding sequence of the sub-blocks and the state change probability of adjacent sub-blocks are statistically analyzed, and the two are combined to form the classification features of the target. Since the distribution probability of the sub-block coding sequence describes the local structure features, and the state change probability of adjacent sub-blocks characterizes the association features between adjacent local structures, the recognition performance of the target is improved. Simulation experiments are carried out on the RCS data of four types of simulation targets, and the experimental results verify the effectiveness of the method. Specific embodiments

[0038] The following combines simulations to prove the effectiveness and progress of the present invention:

[0039] Four types of simulation targets are designed: real target, debris, light decoy, and heavy decoy. The real target is a conical target with geometric dimensions: length 1820 mm, bottom diameter 540 mm; the light decoy is a conical target with geometric dimensions: length 1910 mm, bottom diameter 620 mm; the heavy decoy is a conical target with geometric dimensions: length 600 mm, bottom diameter 200 mm. The precession frequencies of the real target, light decoy, and heavy decoy are 2 Hz, 4 Hz, and 10 Hz respectively. The RCS sequences of the real target, light decoy, and heavy decoy targets are calculated by FEKO, the radar carrier frequency is 3 GHz, and the pulse repetition frequency is 20 Hz. The RCS sequence of the debris is assumed to be a Gaussian random variable with a mean of 0 and a variance of -20 dB. The polarization mode is VV polarization. The calculation target running time is 1400 seconds. The RCS sequence data of each target is divided into 140 frames at intervals of 10 seconds, and the RCS frame data with even frame numbers is taken for training, and the remaining frame data is used as test data. Then, each type of target has 70 test samples.

[0040] Recognition experiments are carried out on four types of targets (real target, debris, light decoy, and heavy decoy) using the comprehensive feature extraction method of the present invention, and the average correct recognition rate reaches about 91%.​

Claims

1. A method for extracting comprehensive features of an RCS sequence, characterized in that, Including the following steps: S1. Define the j-th training RCS data sequence frame of the i-th type of true and false targets obtained as an n-dimensional column vector x ij , where 1 ≤ i ≤ C and 1 ≤ j ≤ N i , and among them, N i is the number of frames of the training RCS sequence of the i-th type of true and false targets, then the total number of frames of the training RCS sequence S2. Perform short-time Fourier transform on the training sample data x of the i-th type of true and false targets ij to obtain the time-frequency spectrogram S ij : S ij = [s ij,km K×M ​ Among them, s ij,km represents the element in the time-frequency spectrogram S ij where k = 1, 2, …, K; m = 1, 2, …, M. k is the row subscript representing the frequency change direction, and m is the column subscript representing the time change direction. K is the number of rows of S ij and M is the number of columns of S ij ; S3. When dividing the sub - blocks of the spectrogram S ij , obtain the sub - block differential binary coding sequence: Perform non-overlapping sliding in the time-frequency spectrogram S with a 4x4 window ij to obtain multiple 4x4 sub-blocks, which form a sub-block set Q ij : Q ij = [P ij,1 P ij,2 …P ij,L ​ P ij,l = [p ij,l,te 4×4 , where t, e = 1, 2, 3, 4​ l＝1,2,…L Among them, P ij,l represents the l-th sub-block matrix in the sub-block set Q ij , p ij,l,te represents an element in the sub-block matrix P ij,l , and L represents the total number of sub-blocks; Calculate the mean value of all elements in the sub-block matrix P ij,l ​ For sub-block P ij,l in p ij,l,te perform the following encoding: Among them, c ij,l,te is the encoding corresponding to p ij,l,te The encodings of the elements in the sub-block form a binary differential encoding sequence c of 0 and 1 in the order of rows ij,l : c ij,l = [c ij,l,11 c ij,l,12 …c ij,l,44 ​ S4. Extract the local structural features of the sub-block differential binary coding distribution: Take the differential coding sequence c ij,l As a binary number, count the number of repeated occurrences of the differential coding sequence values in each sub-block of the entire time-frequency spectrum diagram, and form a vector: Among them, the sub-block differential coding sequence values are sorted in ascending order, the number of these values is L1, π ij,1 is the number of repeated occurrences of the first value, π ij,2 is the number of repeated occurrences of the second value, is the number of repeated occurrences of the L1-th value; Obtain the time-frequency spectrogram S ij The local structure feature vector h of the corresponding sub-block differential binary coding distribution ij : S5. Extract the state change all-local correlation features between adjacent sub-blocks: The distribution interval of the binary coding sequence values of the sub-blocks is evenly divided into 16 sub-intervals, and the central values of each sub-interval are v1, v2, … v 16 , respectively. The number of repetitions of the state change of the binary coding sequence values of adjacent sub-blocks in the time-frequency spectrogram is counted: Among them, indicates that the coding value of the current sub-block is v f (f = 1, 2... 16), the coding value of the adjacent sub-block is v r (r = 1, 2... 16) number of repetitions. After normalizing the statistical matrix by row, the state change feature is obtained: Among them, W ij represents the state change feature matrix; S6. Local structural feature h of the binary coding distribution of the synthesis sub-block ij and the global correlation feature W of the state change between sub-blocks ij , to obtain the synthesis feature vector: [h ij w ij,1 w ij,2 …w ij,16 ​ Among them, w ij,1 , w ij,2 , w ij,16 respectively represent the first row vector, the second row vector, and the sixteenth row vector in matrix W ij .

Images