Aerial target identification and classification method based on SAR image fusion

Abstract

The invention relates to an aerial target identification and classification method based on SAR image fusion, and the method comprises the following steps: S1, obtaining a plurality of SAR images, and carrying out the preprocessing of the SAR images; s2, for the SAR images of the same region, performing fusion based on an energy maximization method of a pixel neighborhood to obtain an SAR fusion image, and finally obtaining an SAR fusion image data set of the target in different postures and region environments; s3, dividing the SAR fusion image data set into a training set and a test set; s4, scattering features are extracted from the SAR fusion images in the training set in the S3, and a corresponding scattering feature training set is established; and S5, taking the scattering feature training set generated in the step S4 as input data, training a random forest model, performing decision classification on an SAR fusion image target, and performing verification on an SAR fusion image test set. According to the method, through target information fusion of multi-band and multi-polarization SAR images, the problems of uncertainty of single information source interpretation and the like are avoided, and effective identification of ground targets in any complex environment can be realized.

Description

Technical Field

[0001] The invention relates to the technical field of electromagnetic scattering modeling, and in particular to an aerial target recognition and classification method based on SAR image fusion. Background Art

[0002] Synthetic aperture radar (SAR) overcomes the influence of weather and lighting conditions on optical remote sensing, so it can achieve all-day and all-weather observation of targets; it can provide detailed mapping data and images under all-weather conditions, and this capability is crucial for ground target identification.

[0003] At the same time, the development of imaging systems towards high resolution, multi-polarization, and multiple working modes has become a trend. Among them, high-frequency band SAR images have a more complete description of the details and appearance of the target, but this type of image has more bright pixels, more obvious clutter, and the target is easily submerged in the surrounding environment; while low-frequency band SAR can penetrate the surface vegetation and find hidden targets that are blocked and buried. Among the different polarization modes in the same band, co-polarization (HH, VV) is more beneficial for finding targets parallel to the radar line of sight, while cross-polarization (HV, VH) is better for detecting targets at a certain angle to the radar line of sight. Experiments show that complete polarization information is conducive to target detection and recognition, and state estimation. Therefore, by fusing multi-band, multi-polarization SAR images and comprehensively utilizing the effective information of multi-band and multi-polarization, it is of great significance for effective target recognition and classification.

[0004] It should be understood that the above statements merely provide background technology related to the present invention and do not necessarily constitute prior art. Summary of the invention

[0005] The purpose of the present invention is to provide an aerial target recognition and classification method based on SAR image fusion, which realizes effective recognition of ground targets in any complex environment by fusing multi-band and multi-polarization SAR images and comprehensively utilizing multi-band and multi-polarization effective information.

[0006] In order to achieve the above-mentioned purpose, the present invention provides an aerial target recognition and classification method based on SAR image fusion, comprising the following steps: S1, obtaining a plurality of SAR images and performing preprocessing; S2, for SAR images of the same area, fusing the SAR fusion image based on the energy summation method of the pixel neighborhood to obtain a SAR fusion image, and finally obtaining a SAR fusion image data set under different target postures and regional environments; S3, dividing the SAR fusion image data set into a training set and a test set; S4, extracting scattering features from the SAR fusion images in the S3 training set, and establishing a corresponding scattering feature training set; S5, using the scattering feature training set generated in S4 as input data, training a random forest model, making decisions and classifying the SAR fusion image targets, and verifying the results on the SAR fusion image test set.

[0007] Preferably, in S1, the preprocessing process is to correct the coordinates of a number of multi-directional and multi-scale SAR images of the same area and unify them into the same coordinate system; and uniformly adjust the above-mentioned several SAR images of the same area into SAR images to be fused with the same pixel size.

[0008] Preferably, the pixel neighborhood energy maximization method utilizes the pixel neighborhood area energy to determine the brightest part in the SAR image.

[0009] Preferably, the method for extracting the energy of the pixel neighborhood is implemented by a sliding neighborhood operation, which specifically includes the following steps: J1. Select a pixel of a SAR image as the central pixel point; J2. Determine the neighborhood range of the central pixel point according to the sliding window; J3. Use a function to calculate the neighborhood area energy of the central pixel point in the neighborhood, and return a scalar value; J4. Slide a sliding window of size 3×3 in the SAR image, stepping one column or one row each time, and calculate the neighborhood area energy of each pixel point in the SAR image; J5. For the pixel points at the edge of the SAR image, consider that the edge is background information and do not calculate the neighborhood area energy; J6. Use a sliding window of the same size to repeat the above steps J1 to J5 for each SAR image, and select the pixels with the largest energy values ​​at the same position in all SAR images to obtain a SAR fusion image.

[0010] Preferably, the neighborhood energy of each pixel is defined as:

[0011]

[0012] Where (x, y) corresponds to the position of the center pixel, E is the i×j size regional window centered at (x, y); D(x+i, y+j) is the wavelet coefficient at the position (x+i, y+j); ω(i, j) is the window weighting coefficient matrix.

[0013] Preferably, the calculation formula of the fused SAR image is:

[0014] F(x,y)=D k (x,y),k=max{E k (x,y)};

[0015] Where F represents the fused image; (x, y) traverses the pixels corresponding to the entire target area; D k (x, y) represents the pixel value corresponding to the original SAR image at the position (x, y); E k (x, y) represents the regional energy based on adjacent pixels corresponding to the original SAR image at the position (x, y); the subscript k represents SAR images of different bands and polarizations.

[0016] Preferably, in S4, the scattering characteristics of the SAR fusion image are calculated by a CLEAN algorithm; the CLEAN algorithm specifically comprises the following steps:

[0017] K1. Generate a two-dimensional radar image based on SAR fusion images; the two-dimensional radar image is composed of a series of finite point sources and their point spread functions; K2. Search for the coordinates and amplitude values ​​of the peak points in the two-dimensional radar image, and extract the strongest peak point scattering features according to the order of intensity {A i ,x i ,y i}; Among them, A i represents the amplitude of the i-th scattering center, (xi,yi) represents the position of the i-th scattering center; K3, subtract the peak point and its sidelobe effect from the i-th two-dimensional radar image to form the i+1-th two-dimensional radar image; K4, determine whether the i+1-th image reaches the preset noise threshold. If so, the algorithm ends and the scattering feature {A i ,x i ,y i}; If not reached, repeat steps K2 to K4 for the image generated by K3.

[0018] Preferably, in S5, the random forest classification algorithm specifically comprises the following steps: L1, model training: using the bootstrap resampling technique to repeatedly randomly extract K training sets from the scattering feature training set with replacement, thereby constructing K CART decision trees accordingly to obtain K classification results; L2, decision classification: the decision results of each decision tree are summarized using the simple majority voting method, and the final classification result is determined by the score formed by the votes of each decision tree, and the specific formula is:

[0019]

[0020] In the formula, H(x) represents the random forest decision result, Y represents the decision possibility, hi(x) represents the i-th decision tree, K represents the number of decision trees, and I(x) represents the decision result of the decision tree.

[0021] In summary, compared with the prior art, the present invention provides an aerial target recognition and classification method based on SAR image fusion. The method analyzes the target pixel intensity in multi-source SAR images, performs pixel-level fusion of target image areas for images of the same area, and introduces a random forest classifier to learn the scattering characteristics of the target to achieve classification and recognition of ground targets. In this method, the uncertainty of single information source interpretation is avoided by fusion of target information of multi-band and multi-polarization SAR images, and the effective recognition of ground targets in any complex environment can be achieved, which greatly expands the scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the algorithm flow chart of the present invention;

[0023] Figure 2 It is a specific flow chart of the CLEAN algorithm of the present invention;

[0024] Figure 3 This is a diagram of the random forest classification method of the present invention. DETAILED DESCRIPTION

[0025] The following is combined with Figures 1 to 3 , the present invention is further explained by describing a preferred specific embodiment in detail.

[0026] It should be noted that the drawings are in a very simplified form and use non-precise proportions. They are only used to conveniently and clearly assist in explaining the embodiments of the present invention, and are not used to limit the conditions for the implementation of the present invention. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention.

[0027] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only the elements explicitly listed, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0028] like Figure 1 As shown, the present invention provides an aerial target recognition and classification method based on SAR image fusion, which specifically comprises the following steps:

[0029] S1. Acquire several synthetic aperture radar (SAR) images and preprocess them;

[0030] S2. For SAR images of the same area, the SAR fusion image is obtained by fusing the SAR images based on the energy of the pixel neighborhood, thereby obtaining a SAR fusion image dataset under different target postures and regional environments;

[0031] S3, dividing the SAR fusion image data set into a training set and a test set;

[0032] S4, extracting scattering features from the SAR fusion image in the S3 training set, and establishing a corresponding scattering feature training set;

[0033] S5. Using the scattering feature training set generated in S4 as input data, training a random forest model, making decisions and classifying SAR fusion image targets, and verifying the results on a SAR fusion image test set.

[0034] Among them, in the S1, the preprocessing process is to correct the coordinates of several multi-directional and multi-scale SAR images of the same area and unify them into the same coordinate system; at the same time, the several SAR images of the same area are uniformly adjusted to the SAR images to be fused with the same pixel size.

[0035] Among them, in the above S2, the energy of the pixel neighborhood is taken as the maximum, and the energy of the pixel neighborhood area is used to determine the brightest part of the SAR image. Specifically, for several SAR images to be fused in the same area, the neighborhood area energy of each pixel in one of the SAR images is first calculated, and the energy of the neighborhood area of ​​the corresponding pixel point in all other SAR images is compared. When the neighborhood area energy of a certain pixel point in a certain SAR image is greater than the neighborhood area energy of the pixel point at the corresponding position in all other SAR images, the grayscale value of the corresponding pixel point in the SAR image is used as the grayscale value of the pixel point at the corresponding position in the SAR fusion image. This fusion method has the characteristics of small amount of calculation data, strong real-time performance, and is easy to implement.

[0036] Furthermore, the energy of the pixel neighborhood is maximized by a sliding neighborhood operation, and the energy of the neighborhood area of ​​each pixel in each SAR image is calculated by the sliding neighborhood operation. It should be noted that the neighborhood of a pixel is a series of pixels related to its central pixel, which is a rectangular block. When moving from one pixel to the next pixel on the image, the sliding block slides in the same direction. Therefore, the sliding neighborhood operation includes the following steps:

[0037] J1. Select a pixel of a SAR image as the central pixel;

[0038] J2, determine the neighborhood range of the central pixel point according to the sliding window;

[0039] J3. Use the function to calculate the neighborhood area energy of the central pixel in the neighborhood and return a scalar value;

[0040] J4. Slide the sliding window (the window size is 3×3) in the SAR image, stepping one column or one row at a time, and calculate the neighborhood area energy of each pixel in the SAR image;

[0041] J5. For the pixel points at the edge of the SAR image, the edge is considered as background information and the energy of its neighborhood area is not calculated;

[0042] J6. Use a sliding window of the same size to repeat the above steps J1 to J5 for each SAR image, and select the pixels with the largest energy value at the same position in all SAR images to obtain a SAR fusion image.

[0043] Among them, the neighborhood area energy of each pixel is defined as shown in the following formula:

[0044]

[0045]

[0046] Where (x, y) corresponds to the position of the center pixel, E is the i×j size regional window centered at (x, y); D(x+i, y+j) is the wavelet coefficient at the position (x+i, y+j); ω(i, j) is the window weighting coefficient matrix.

[0047] Furthermore, assuming that the fused SAR image is F, the formula is:

[0048] F(x,y)=D k (x,y),k=max{E k (x,y)},

[0049] Where F represents the fused image; (x, y) traverses the pixels corresponding to the entire target area; D k (x, y) represents the pixel value corresponding to the original SAR image at the position (x, y); E k (x, y) represents the regional energy based on adjacent pixels corresponding to the original SAR image at the position (x, y); the subscript k represents SAR images of different bands and polarizations.

[0050] Wherein, in said S3, the fused SAR image data set is randomly divided, wherein the training set accounts for 70% to 90% of the total data set, and correspondingly, the test set accounts for 10% to 30% of the total data set.

[0051] For targets in SAR images, some scatterers with simple geometric structures (such as points, corners, edges, etc.) on the object will produce strong scattered echoes. These strong scatterers are called scattering centers. There is a clear difference between the SAR image and the background, and the scattering center is manifested as a cluster of strong scattering points. Therefore, under different radar illumination angles, the position of the scattering center often does not change much, and it is relatively stable information, so the target can be reconstructed based on the scattering center.

[0052] Assume that the target is composed of N scattering centers and the incident field is a plane wave along the zr direction. Then the scattering field of a scattering center of the target can be expressed as follows:

[0053]

[0054] In the formula, Indicates along The scattered field in the direction, c is the speed of light, f is the frequency, and A represents the amplitude.

[0055] Therefore, the normalized scattering field obtained by adding up the echoes of N scattering centers can be expressed as:

[0056]

[0057] Where N is the number of scattering centers, A i represents the amplitude of the i-th scattering center, c is the speed of light, f is the frequency, (x i ,y i ) represents the position of the i-th scattering center.

[0058] It should be noted that the SAR image can be represented as a series of scattered finite point sources and point spread functions, namely:

[0059]

[0060] Among them, h(x,)y is the point spread function, and its physical meaning is the impulse response function of the inverse synthetic aperture radar (Inverse-Synthetic-Aperture-Rada, ISAR) imaging system. Under the condition of small-angle imaging, the expression of the point spread function is a two-dimensional sinc function. In addition, it should be noted that another explanation of the point spread function is that the actual acquisition process of the scattering field data of the radar target is within a limited bandwidth and angle range, and the ideal Fourier transform integral limit is infinite, so it is impossible to image the ideal shape of the point source, and the target image has certain side lobes, and spectrum energy leakage occurs. Too high side lobes will cause nearby weak targets to be submerged, and it is difficult to extract the submerged weak scattering center directly from the ISAR image. Therefore, in the preferred embodiment of the present invention, the CLEAN algorithm is used to eliminate the influence of the side lobes.

[0061] Further, in said S4, the scattering characteristics of said SAR fusion image {A i ,x i ,y i} can be extracted by the CLEAN algorithm, where A i represents the amplitude of the i-th scattering center, (xi,yi) represents the position of the i-th scattering center; the specific process of the CLEAN algorithm is as follows Figure 2 As shown, the following steps are included:

[0062] K1. Generate a two-dimensional radar image based on the SAR fusion image; wherein the two-dimensional radar image is composed of a series of finite point sources and their point spread functions;

[0063] K2, search for the coordinates and amplitude values ​​of the peak points in the two-dimensional radar image, and extract the strongest peak point scattering features according to the order of intensity {A i ,x i ,y i};

[0064] K3, subtracting the peak point and its sidelobe effect from the i-th two-dimensional radar image to form the i+1-th two-dimensional radar image;

[0065] K4, determine whether the i+1th image reaches the preset noise threshold. If so, the algorithm ends and the scattering feature {A i ,x i ,y i}; If not reached, repeat steps K2 to K4 for the image generated by K3.

[0066] Among them, in the S5, as Figure 3As shown, the random forest algorithm is used for classification and recognition, and the scattering feature training set generated by S4 is used as input data. The random forest classification algorithm can be summarized as the following steps:

[0067] L1, model training; use the bootstrap resampling technique (Bootstrap sampling) to repeatedly randomly extract K training sets from the scattering feature training set with replacement, thereby constructing K CART decision trees accordingly and obtaining K classification results.

[0068] L2, decision classification; the decision results of each decision tree are summarized using the simple majority voting method, and the final classification result is determined by the scores formed by the votes of each decision tree. The specific formula is:

[0069]

[0070] In the formula, H(x) represents the random forest decision result, Y represents the decision possibility, hi(x) represents the i-th decision tree, K represents the number of decision trees, and I(x) represents the decision result of the decision tree.

[0071] In summary, the present invention provides an aerial target recognition and classification method based on SAR image fusion. The method analyzes the target pixel intensity in multi-source SAR images, performs pixel-level fusion of target image areas for images of the same area, and introduces a random forest classifier to learn the scattering characteristics of the target to achieve classification and recognition of ground targets. In addition, in the method, the uncertainty of single information source interpretation is avoided by fusion of target information of multi-band and multi-polarization SAR images, and the effective recognition of ground targets in any complex environment can be achieved, which greatly expands the scope of application.

[0072] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for aerial target recognition and classification based on SAR image fusion, characterized in that: The following steps are involved: S1. Acquire several SAR images and perform preprocessing; S2, for the SAR images of the same area, the energy of the pixel neighborhood is combined to obtain SAR fusion images, finally obtaining SAR fusion image datasets in different target postures and regional environments; S3, dividing the SAR fusion image data set into a training set and a test set; S4, extracting scattering features from the SAR fusion image in the S3 training set, and establishing a corresponding scattering feature training set; S5. Using the scattering feature training set generated in S4 as input data, training a random forest model, making decisions and classifying SAR fusion image targets, and verifying the results on a SAR fusion image test set.

2. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 1, characterized in that: In the above S1, the preprocessing process is to correct the coordinates of a number of multi-directional and multi-scale SAR images of the same area and unify them into the same coordinate system; And the above-mentioned several SAR images of the same area are uniformly adjusted to SAR images to be fused with the same pixel size.

3. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 1, characterized in that: The pixel neighborhood energy maximization method uses the pixel neighborhood area energy to determine the brightest part in the SAR image.

4. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 3, characterized in that: The energy maximization method of the pixel neighborhood is implemented by a sliding neighborhood operation, which specifically includes the following steps: J1. Select a pixel of a SAR image as the central pixel; J2, determine the neighborhood range of the central pixel point according to the sliding window; J3. Use the function to calculate the neighborhood area energy of the central pixel in the neighborhood and return a scalar value; J4. Slide a 3×3 sliding window in the SAR image, stepping one column or one row at a time, and calculate the neighborhood area energy of each pixel in the SAR image; J5. For the pixel points at the edge of the SAR image, the edge is considered as background information and the energy of its neighborhood area is not calculated; J6. Use a sliding window of the same size to repeat the above steps J1 to J5 for each SAR image, and select the pixels with the largest energy value at the same position in all SAR images to obtain a SAR fusion image.

5. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 4, characterized in that: The energy definition formula of the neighborhood area of ​​each pixel is: Where (x, y) corresponds to the position of the center pixel, E is the i×j size regional window centered at (x, y); D(x+i, y+j) is the wavelet coefficient at the position (x+i, y+j); ω(i, j) is the window weighting coefficient matrix.

6. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 5, characterized in that: The calculation formula of the fused SAR image is: F(x,y)=D k (x,y),k=max{E k (x,y)}; Where F represents the fused image; (x, y) traverses the pixels corresponding to the entire target area; D k (x, y) represents the pixel value corresponding to the original SAR image at the position (x, y); E k (x, y) represents the regional energy based on adjacent pixels corresponding to the original SAR image at the position (x, y); the subscript k represents SAR images of different bands and polarizations.

7. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 1, characterized in that: In the S4, the scattering characteristics of the SAR fusion image are calculated by the CLEAN algorithm; The CLEAN algorithm specifically comprises the following steps: K1. Generate a two-dimensional radar image based on the SAR fusion image; wherein the two-dimensional radar image is composed of a series of finite point sources and their point spread functions; K2, search for the coordinates and amplitude values ​​of the peak points in the two-dimensional radar image, and extract the strongest peak point scattering features {Ai,xi,yi} according to the order of intensity; where A i represents the amplitude of the i-th scattering center, (x i ,y i ) represents the position of the i-th scattering center; K3, subtracting the peak point and its sidelobe effect from the i-th two-dimensional radar image to form the i+1-th two-dimensional radar image; K4, determine whether the i+1th image reaches the preset noise threshold. If so, the algorithm ends and the scattering feature {A i ,x i ,y i }; If not reached, repeat steps K2 to K4 for the image generated by K3.

8. The aerial target recognition and classification method based on SAR image fusion as claimed in claim 1, characterized in that: In the above S5, the random forest classification algorithm specifically comprises the following steps: L1, model training: Use the bootstrap resampling technique to repeatedly randomly extract K training sets from the scattering feature training set with replacement, thereby constructing K CART decision trees accordingly and obtaining K classification results; L2, decision classification: The decision results of each decision tree are summarized using the simple majority voting method, and the final classification result is determined by the score formed by the votes of each decision tree. The specific formula is: In the formula, H(x) represents the random forest decision result, Y represents the decision possibility, hi(x) represents the i-th decision tree, K represents the number of decision trees, and I(x) represents the decision result of the decision tree.

Images