CN112836707B

CN112836707B - ISAR image aerial target length feature extraction method

Info

Publication number: CN112836707B
Application number: CN202110031079.4A
Authority: CN
Inventors: 王鹏辉; 辛萌; 邵帅; 刘宏伟; 丁军; 陈渤
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2021-01-11
Filing date: 2021-01-11
Publication date: 2022-03-22
Anticipated expiration: 2041-01-11
Also published as: CN112836707A

Abstract

The invention provides a method for extracting ISAR image aerial target length features, which can be used for radar aerial target identification. The method comprises the following implementation steps: 1) preprocessing an ISAR image; 2) obtaining a region of interest ROI image with the same position dimension and distance dimension resolution; 3) acquiring a binary image of an ROI (region of interest) image; 4) acquiring a binary image after the horizontal stripe noise of the binary image is suppressed; 5) and obtaining an extraction result of the length characteristics of the aerial target in the binary image. According to the method, smooth filtering is carried out on the ROI image of the region of interest by adopting a Gaussian filtering method, so that the target in the image is smooth, continuous and complete, the real shape information of the target can be well kept, and the accuracy of the ISAR image in-air target length feature extraction is effectively improved.

Description

ISAR image aerial target length feature extraction method

Technical Field

The invention belongs to the field of radar image processing, and relates to an ISAR image aerial target length feature extraction method, which can realize the extraction of aerial target length features and can be used for radar aerial target identification.

Background

Target identification is mainly divided into optical means and radar means. The optical means is a traditional detection means, the technology is mature, and the target is monitored mainly by the reflected light of the target, so that the optical means is greatly influenced by day and night and weather. Compared with an optical means, the radar means actively transmits electromagnetic waves to the space and utilizes received echoes to complete the monitoring of the target, so that the influence of day and night and weather environments can be effectively overcome, the radar has the characteristics of monitoring the target continuously all day long, all day long and 24 hours, and along with the development of radar technology, the radar can not only detect, track and position the target, but also realize high-resolution imaging of the target. Radar imaging techniques were proposed as early as the 50 s of the 20 th century, and can be classified according to the imaging principles: synthetic Aperture Radar (SAR) and Inverse Synthetic Aperture Radar (ISAR). The inverse synthetic aperture radar imaging is that the radar is kept still, two-dimensional imaging of moving targets such as airplanes, missiles, satellites and ships can be achieved by emitting broadband electromagnetic waves and utilizing relative motion of the targets and the radar to obtain ISAR images of the targets, detailed information such as the size, shape, structure and posture of the targets can be described through the two-dimensional ISAR images, and the attributes, categories or types of the targets can be distinguished by utilizing the information, so that powerful support is provided for radar target feature extraction and classification and recognition.

The overall dimension characteristics of the target, such as target length, area, contour and the like, have intuitiveness, resolution, easy extraction and definite physical significance, so the method is widely used for extracting the target characteristics of the ISAR image, wherein the length characteristics of the target are one of effective characteristics of target classification and identification, the accuracy of the target length characteristic extraction result is an important index for judging the target length characteristic extraction method, and the higher the accuracy is, the better the performance of the method is.

In the prior art, an ISAR image target length feature extraction method mainly aims at a ship target, extracts the ship target in an image by performing speckle noise suppression, cross-stripe interference suppression, morphological filtering and image segmentation operations on an ISAR image, and then extracts the length feature of the ship target by using a Principal Component Analysis (PCA) method. The method fills and connects ship targets in the image by using a morphological filtering method, improves the connectivity of the targets in the image, and enables the targets in the image to be continuous and complete.

Compared with ship targets, scattering points in the ISAR images of the aerial targets are more sparse and discrete, so when the method is used for extracting the length features of the aerial targets, the aerial targets in the images can be fully filled and connected only by adopting larger structural elements for morphological filtering, but the aerial targets in the images are excessively connected and filled due to the overlarge structural elements, the aerial targets in the images are distorted, and the accuracy of extracting the length features of the aerial targets in the ISAR images can be seriously influenced.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides an ISAR image aerial target length feature extraction method which is used for solving the problem that the ISAR image aerial target length feature extraction result in the prior art is low in accuracy.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:

(1) preprocessing an ISAR image:

(1a) obtaining an ISAR image A containing an aerial target with the size of m multiplied by n, wherein the resolution of a distance dimension corresponding to the row dimension of the A is P_rThe resolution of the azimuth dimension corresponding to the column dimension is P_aObtaining the amplitude image A' of A by taking the modulus value of the original complex data of A, wherein m is more than or equal to 100, and n is more than or equal to 100;

(1b) normalizing the amplitude image A ' of the A to obtain a gray image I with the pixel amplitude of 0-255, and extracting a region of interest ROI image I ' with the size of m ' × n ' in the image I, wherein m ' is more than or equal to 30 and less than or equal to m ', and n ' is more than or equal to 30 and less than or equal to n;

(2) obtaining a ROI image I with equal position dimension and distance dimension resolution_A：

Calculating P_aAnd P_rRatio of

When k is more than or equal to 1, carrying out pixel interpolation operation on the ROI image I ' of the region of interest in the azimuth dimension, otherwise, carrying out pixel extraction operation on the ROI image I ' of the region of interest in the azimuth dimension to obtain the ROI image I of the region of interest with the equal azimuth dimension resolution and distance dimension resolution and with the size of m ' × n_AWherein n ═ k × n']，[·]Represents a rounding operation;

(3) obtaining ROI image I_ABinary image I of_C：

(3a) ROI image I of region of interest by Gaussian filtering method_ACarrying out smooth filtering to obtain a filtered image I_BAnd to I_BBackup is carried out to obtain I_BBackup image I of_B′；

(3b) For backup image I_B' performing threshold segmentation to obtain a threshold-segmented binary image I_BAAnd to I_BABackup is carried out to obtain I_BABackup image I of_BA′；

(3c) For backup image I_BA' negation is performed and the filtered image I is masked with the negation result as a mask matrix_BCarrying out masking operation to obtain a masked image I_BB；

(3d) To I_BBPerforming threshold segmentation to obtain a threshold-segmented binary image I_BCAnd for the binary image I_BAAnd a binary image I_BCPerforming logical OR operation to obtain ROI image I_ABinary image I of_C；

(4) Obtaining a binary image I_CBinary image I after horizontal stripe noise suppression_D：

(4a) Initializing the iteration times as T, the maximum iteration times as T, T is more than or equal to 1, and obtaining a binary image I_CThe binary image after the t-th horizontal stripe noise suppression is I_D ^tAnd let t equal to 1, I_D ^t＝I_C；

(4b) Statistical binary image I_D ^tThe pixel number of each row of the pixel with the pixel amplitude value of 1 is obtained, and a pixel number sequence N is obtained as { N ═ N₁，n₂，…，n_i，…，n_m′And judging whether the maximum value MAX and N' in N satisfy

If yes, executing step (4c), otherwise, executing step I_D ^tAs a binary image I_CBinary image I subjected to horizontal stripe noise suppression_DAnd output, wherein n_iRepresenting a binary image I_D ^tThe number of pixels in the ith row having a pixel amplitude of 1;

(4c) carrying out first-order forward difference on the pixel number sequence N to obtain a first-order forward difference sequenceColumn Δ N ═ Δ N₁，Δn₂，…Δn_d，…，Δn_m′-1And according to the position d corresponding to the maximum value in the delta N₁Position d corresponding to the minimum value₂Determining a binary image I_D ^tLine range d where middle horizontal stripe noise is located₁+1～d₂And the width W ═ d of the horizontal streak noise₂-d₁Wherein, Δ n_dRepresenting the difference, Δ n, between the number of pixels of adjacent rows having a pixel amplitude of 1_d＝n_d+1-n_d，0＜d₁≤m′-1,0＜d₂≤m′；

(4d) Selecting a binary image I_D ^tBinary image I with medium size of l × n ″_CAAnd count of I_CAThe number of pixels with the pixel amplitude of 1 in each row is obtained, and a pixel number sequence V ═ V is obtained₁，V₂，…V_j，…，V_n″Where l ═ l₂-l₁+1，l₁And l₂Respectively representing binary images I_D ^tStarting and ending lines selected when selection is performed,/₁＝max(1,d₁+1-W)，l₂＝min(m′,d₂+ W), max (,) indicates the maximum value, min (,) indicates the minimum value, V_jRepresenting a binary image I_CAThe number of pixels with the pixel amplitude value of 1 in the jth row of (1), j is more than 0 and less than or equal to n';

(4e) when V is_jWhen the image is less than or equal to W, the binary image I is processed_D ^tD (d) of₁+1 line to d₂The amplitude value of the jth column pixel in the row is set to be 0, otherwise, the binary image I is kept_D ^tD (d) of₁+1 line to d₂The amplitude of the jth column pixel in the row is unchanged, and the binary image I after the t-th horizontal stripe noise suppression is obtained_D ^t；

(4f) Judging whether T is greater than or equal to T, if so, obtaining a binary image I_CBinary image I after horizontal stripe noise suppression_DAnd outputting, otherwise, making t equal to t +1, and executing the step (4 b);

(5) obtaining a binary image I_DThe extraction result of the length characteristic of the aerial target in (1):

(5a) extracting a binary image I_DObtaining the coordinate information of the pixel with the middle amplitude value of 1 to obtain a two-dimensional coordinate matrix

And carrying out zero equalization on the G to obtain a zero-equalized two-dimensional coordinate matrix

Then adopting Principal Component Analysis (PCA) to pair G₀Linear transformation is carried out to obtain a two-dimensional coordinate matrix after linear transformation

Wherein (x)_q,y_q) Representing a binary image I_DIs located at x_qLine y_qCoordinates of pixels of the column, (x)_q′,y_q') represents (x)_q,y_q) Zero-averaged coordinates, (xs)_q,ys_q) Represents (x)_q′,y_q') coordinates after linear transformation, h denotes the binary image I_DTotal number of pixels of medium amplitude 1 [ ·]^TRepresenting a transpose operation;

(5b) according to G_SDetermining a line number q1 corresponding to the minimum value and a line number q2 corresponding to the maximum value of the 1 st column element in the binary image I_DHead coordinates (x) of hollow target_q1,y_q1) And tail coordinates (x)_q2,y_q2) And combining the distance dimension resolution P in the step (1)_rCalculating a binary image I_DLength of target in hollow:

compared with the prior art, the invention has the following advantages:

1. according to the method, the obtained ROI image of the region of interest with the equal position dimension and distance dimension resolution is subjected to smooth filtering by adopting a Gaussian filtering method, so that the smoothness, continuity and integrity of the hollow target in the image are improved, the real shape information of the hollow target can be well maintained, the shape of the hollow target in the binary image of the ROI image of the region of interest obtained by image segmentation is closer to the real shape of the hollow target, the problem that the hollow target in the image is distorted when the ROI image of the region of interest with the equal position dimension and distance dimension resolution is subjected to morphological filtering in the prior art is solved, and the extracted head coordinate and tail coordinate of the hollow target can be ensured to be more accurate. Simulation results show that compared with the prior art, the method effectively improves the accuracy of extracting the aerial target features of the ISAR images.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a magnitude image of an ISAR image used in an embodiment of the present invention;

FIG. 3 is an ROI image extracted in an embodiment of the present invention;

FIG. 4 is a ROI image with equal resolution in the azimuth dimension and the distance dimension acquired in the embodiment of the present invention;

FIG. 5 is an ROI image I of an embodiment of the present invention_AFiltered image I obtained by smoothing filtering_B；

FIG. 6 shows the pair I in the embodiment of the present invention_BBackup image I of_B' binary image I obtained by performing threshold segmentation_BA；

FIG. 7 shows a pair I in an embodiment of the present invention_BBPerforming threshold segmentation to obtain binary image I_BC；

FIG. 8 is a ROI image I obtained in an embodiment of the present invention_ABinary image I of_C；

FIG. 9 is a diagram of a binary image I according to an embodiment of the present invention_CBinary image I obtained by performing horizontal stripe noise suppression_D；

Fig. 10 is a comparison graph of simulation results of the present invention and the prior art for the accuracy of extracting the target length feature in the air of the ISAR image used in the present embodiment.

Detailed Description

The invention is described in further detail below with reference to the figures and specific examples.

Referring to fig. 1, the present invention includes the steps of:

step 1) carrying out preprocessing on an ISAR image:

(1a) obtaining an ISAR image A containing an aerial target with the size of m multiplied by n, wherein the resolution of a distance dimension corresponding to the row dimension of the A is P_rThe resolution of the azimuth dimension corresponding to the column dimension is P_aThe original image data of the ISAR image A is complex data, complex information on each pixel is the expression of the reflectivity of a corresponding scattering point on an aerial target and a surrounding background, and the complex data is difficult to process and low in efficiency, so that a modulus value is taken for the original complex data of the ISAR image A to obtain an amplitude image A 'of the A, the data of the amplitude image A' is real data and is convenient for subsequent processing, wherein m is more than or equal to 100, and n is more than or equal to 100; in this example, the selected aerial target is an airplane target, that is, 1 ISAR image a of the airplane target in the external field test is selected, and the magnitude image a' of the ISAR image a is shown in fig. 2, and has a size of 512 × 256 and a distance dimension resolution P_r0.3000m, azimuth dimension resolution P_a＝0.5390m；

(1b) Normalizing the amplitude image A' to obtain a gray level image I with the pixel amplitude of 0-255, wherein the normalization is performed according to the following formula:

wherein, a ' (x, y) is the amplitude value of the pixel in the x-th row and the y-th column in the amplitude image a ', I (x, y) is the amplitude value of the pixel in the x-th row and the y-th column in the grayscale image I obtained after normalization, and cmax is the maximum amplitude value of the pixel in the amplitude image a ';

compared with an optical image and an SAR image, the background and the target structure in the ISAR image are relatively single and simple, and the situations of complex texture of a target, disordered background and the like do not exist, so that the normally obtained gray level image I only consists of two regions, one region is a region consisting of pixel points on an airplane target, is called a foreground region of the gray level image I and is also called an airplane target region, the other region is a region consisting of pixel points on a non-airplane target, is called a background region of the gray level image I, the background region of the gray level image I is usually larger than the foreground region, if the gray level image I is processed, a large number of pixel points in the background region do not contribute to subsequent target length characteristic extraction and cause calculation efficiency reduction, therefore, a region of interest ROI image I' in the gray level image I needs to be extracted, the region of interest refers to an approximate minimum region containing a research target region in the processed image, the shape of the region can be rectangular, circular, elliptical, irregular polygon, etc., in this example, a rectangular region is selected, and can be directly used as an image for subsequent image processing, and the specific step of extracting the ROI image I ' with the size of m ' × n ' in the gray image I is as follows:

(1b1) calculating the sum of the amplitudes of all pixels in each row in the gray level image I to obtain a row pixel amplitude sum sequence RS [ RS ]₁,rs₂,…,rs_x,…,rs_m]And calculating the mean value rm of the elements in RS, wherein RS_xThe sum of all pixel amplitudes of the x-th row representing the gray-scale image I;

(1b2) calculating each element RS in RS_xDifference from rm gives the sequence R ═ R₁,r₂,…,r_x,…r_m]And the minimum value x of the positions corresponding to the elements larger than 0 in the R is determined_minAnd maximum value x_maxRespectively as the start and end lines of a region of interest of the gray image I, where r_x＝rs_x-rm；

(1b3) Calculating the sum of the amplitudes of all pixels in each column in the gray-scale image I to obtain a column pixel amplitude sum sequence CS [ CS ]₁,cs₂,…,cs_x,…,cs_m]And calculating the mean value cm of elements in CS, wherein CS_yThe sum of all pixel amplitudes of the y-th column representing the gray-scale image I;

(1b4) computing each element CS in the CS_yDifference from cm gives the sequence C ═ C₁,c₂,…,c_x,…c_m]And the minimum value y of the positions corresponding to the elements larger than 0 in C_minAnd maximum value y_maxRespectively as the start column and the end column of the region of interest of the gray image I, wherein c_y＝cs_y-cm；

(1b5) Will be the xth of the gray image I_minGo to the x_maxLine, and y_minColumn to y_maxThe rectangular area image formed by the columns and having a size of m '× n' is determined as the region of interest ROI image I 'of the gray image I, where m' ═ x_max-x_min，n′＝y_max-y_min；

In this example, the obtained ROI image I 'in the gray image I is shown in fig. 3, and the size of the ROI image I' is 230 × 180, and as can be seen from fig. 3, the ROI image is an image of an approximately smallest rectangular region containing the aircraft target, wherein the ROI image contains both the complete aircraft target region image and the background region as little as possible, and the computational complexity is reduced as much as possible without affecting the target feature extraction result.

Step 2) obtaining a ROI image I of the region of interest with equal position dimension and distance dimension resolution_A：

Since the ISAR image A in step (1a) has a distance dimension resolution depending on the bandwidth of the measuring radar and an orientation dimension resolution depending on the rotation speed of the non-cooperative target, the ISAR image A has a problem that the distance dimension resolution is not equal to the orientation dimension resolution, since the distance dimension resolution of the ROI image of the region of interest is equal to the distance dimension resolution P of the ISAR image A_rThe azimuth dimension resolution is equal to the azimuth dimension resolution P of the ISAR image A_aTherefore, the ROI image of the region of interest also has the problem that the distance dimension resolution and the azimuth dimension resolution are not equal, when the ROI image I of the region of interest_AWhen the azimuth dimension resolution is not equal to the distance dimension resolution, the head coordinate and the tail coordinate of the aerial target obtained subsequently have a large deviation from the real head coordinate and the real tail coordinate, so that a large error exists in the result of the length feature extraction of the aerial target, and therefore an image with the azimuth dimension resolution equal to the distance dimension resolution needs to be obtained. Distance dimension of ISAR image obtained when same radar system images aerial targetThe resolutions are equal, the azimuth dimension resolutions are usually different from each other, so that the image with the azimuth dimension and the distance dimension resolutions equal is obtained by adjusting the azimuth dimension resolution to be equal to the distance dimension resolution, and the specific acquisition step is as follows: calculating the azimuthal dimension resolution P_aResolution in the dimension of distance P_rRatio of

When k is larger than or equal to 1, adopting a nearest neighbor interpolation method to carry out pixel interpolation operation on the ROI image I ' of the region of interest in an azimuth dimension to enable the resolution of the azimuth dimension to be equal to the resolution of a distance dimension, otherwise, adopting the nearest neighbor interpolation method to carry out pixel extraction operation on the ROI image I ' of the region of interest in the azimuth dimension to enable the resolution of the azimuth dimension to be equal to the resolution of the distance dimension, and obtaining the ROI image I of the region of interest with the equal resolution of the azimuth dimension and the resolution of the distance dimension and with the size of m ' × n_AWherein n ═ k × n']，[·]Represents a rounding operation; the distance dimensional resolution of the region of interest ROI image in this example is equal to the distance dimensional resolution P of the original ISAR image A_rThe azimuth dimension resolution is equal to the azimuth dimension resolution P of the original ISAR image A_a. The common interpolation method comprises a nearest neighbor interpolation method, a bilinear interpolation method and a bicubic interpolation method, the common extraction method comprises the nearest neighbor interpolation method, the bilinear interpolation method or the bicubic interpolation method, the nearest neighbor interpolation method is small in calculation amount and easy to achieve, the nearest neighbor interpolation method is used for conducting pixel interpolation operation on the ROI image of the region of interest in the azimuth dimension, and the nearest neighbor interpolation method is also used for conducting pixel extraction operation on the ROI image of the region of interest in the azimuth dimension. In this example, the azimuthal dimension resolution P is calculated_aResolution in the dimension of distance P_rRatio of

Therefore, the nearest neighbor interpolation method is adopted to carry out pixel interpolation operation on the ROI image I' in the azimuth dimension to obtain the ROI image I with equal resolution of the azimuth dimension and the distance dimension_AAs shown in FIG. 4, I_AHas a size of 230 × 230, and has an azimuth dimension resolution and a distance dimension resolution of 0.3000m。

Step 3) obtaining ROI image I_ABinary image I of_C：

Since the length feature extraction of the aerial target is based on the ROI image I_AThe method is realized by the characteristics of a foreground region, namely an aerial target, so that firstly, the target in the ISAR image needs to be extracted from the background, namely, the image is segmented, and the segmentation effect directly determines the accuracy of feature extraction, so that the image segmentation is a key step in the feature extraction, and particularly, an ROI image I of a region of interest is obtained_ABinary image I of_CComprises the following steps:

(3a) ROI image I of region of interest by Gaussian filtering method_ACarrying out smooth filtering, wherein the size of a window of Gaussian filtering is 5 multiplied by 5-9 multiplied by 9, and obtaining a filtered image I_BAnd to I_BBackup is carried out to obtain I_BBackup image I of_B′；

Due to the fluctuation characteristic of the amplitude of the target echo, the targets in the obtained ISAR image are distributed in sparse, discrete and isolated scattering points, and the ROI image I of the region of interest_AThe same property is also exhibited by the image in (1), as shown in FIG. 4_AThe scattering points of the target in (1) are sparse, discrete and isolated, which often results in the failure to match I_APerforming effective division if directly to I_AAnd (3) performing image segmentation, wherein the edge contour of the target in the obtained binary image is neither complete nor continuous, so that the head and tail coordinates of the subsequently obtained target have large errors, and the accuracy of the ISAR image aerial target length feature extraction is seriously influenced. Therefore, before image segmentation, ROI image I is needed_AThe target areas in the image are connected and filled, so that the connectivity of the target in the image is improved, and the target is smooth, continuous and complete.

A commonly used method for image filling, connecting and deleting is morphological filtering, and although this method can connect and fill a target region in an image, since scattering points in ISAR images of airborne targets such as ship targets and airplanes are more sparse and discrete than scattering points in the images, a region-of-interest ROI image I is obtained_AImage of (1)Also presents the same property, adopts a morphological filtering method to I_AWhen filling and connecting the aerial targets in the middle, the I can be filled and connected by adopting larger structural elements_AThe aerial target in (1) achieves sufficient filling and connection, but the I can be caused by overlarge structural elements_AThe aerial target in the ISAR image is excessively connected and filled, the edge of the aerial target is enlarged, the aerial target in the image is distorted, and the accuracy of the length feature extraction of the aerial target in the ISAR image is seriously influenced. Aiming at the defects of morphological filtering, the method provides a Gaussian filtering method for a region of interest ROI image I_AAnd performing smooth filtering.

The Gaussian filter is a linear smoothing filter, is suitable for eliminating Gaussian noise, is most widely applied to the field of image processing and is used for optical image denoising, generally, the noise is randomly distributed in an optical image, the amplitude of a noise point is far deviated from the average value of pixel amplitudes in a neighborhood, namely, the amplitude of the noise point is mutated relative to the pixel amplitude in the neighborhood, so that the weighted average of the pixel amplitudes in the neighborhood is used for replacing the amplitude of the noise point, the pixel amplitude mutation is inhibited, the amplitude of the noise point is made to obey the statistical characteristics of pixel points in the neighborhood, and the optical image denoising is realized. However, compared with the optical image, the ISAR image usually shows a sparse and isolated scattering center distribution form, the missing pixels on the target in the ISAR image, i.e. the pixels with smaller amplitudes on the target, are essentially scattering points with weaker echoes on the target, usually the amplitudes of the missing pixels on the target are obviously smaller than the amplitudes of the pixels in the neighborhood, i.e. the amplitudes of the pixels in the neighborhood are mutated, so that the missing pixels on the target in the ISAR image have the same characteristics as the noise in the optical image, and thus the missing pixels on the target in the ISAR image can be regarded as "noise", the "noise" in the image is suppressed by performing smooth filtering on the ISAR image, and essentially the missing pixels on the target are filled by using the weighted average amplitudes of the pixels in the neighborhood to improve the amplitudes of the missing pixels on the target so as to make the missing pixels obey the statistical characteristics of the pixels in the neighborhood, smoothing, continuing and completing the target in the image; and because of the Gaussian filtering tool in the smooth filteringThe method has smoothing effect and can better retain the edge and shape information of the target, so that the method provides a Gaussian filtering method for the ROI image I_AThe smooth filtering is carried out, so that the aerial target in the image is more complete, the boundary is clear and continuous, the real edge and shape information of the aerial target can be better kept, the problem of aerial target distortion caused by a morphological filtering method is solved, the shape of the aerial target in the binary image of the ROI image of the region of interest obtained by image segmentation is closer to the real shape of the aerial target, and the extraction result of the aerial target length feature is more accurate.

The window size of the gaussian filter is selected by noting that: the smaller the window is, the worse the connecting and filling effects on the target in the image are, the larger the filtering template is, the worse the edge retaining effect on the target in the image is, and the blurriness of the target boundary is, so the window size of the Gaussian filtering is generally 5 × 5-9 × 9, the window size of the Gaussian filtering in the example is 7 × 7, the standard deviation is 20, and the obtained filtered image I_BAs shown in FIG. 5, it can be seen that compared to image I before filtering_A，I_BThe target of the airplane is smoother, continuous and complete, and meanwhile, the real shape information of the aerial target is well kept;

(3b) to I_BBackup image I of_B' threshold segmentation is performed by using the OSTU method, the segmentation threshold is T1, and I is subjected to threshold T1_B' binarization is carried out to obtain a binary image I_BAAnd to I_BABackup is carried out to obtain I_BABackup image I of_BA', the binary image I obtained in this example_BAAs shown in fig. 6.

Referring to fig. 6, a set of pixel points with an amplitude of 1 in the binary image is referred to as a foreground region of the binary image, and a set of pixel points with an amplitude of 0 is referred to as a background region, where a segmented binary image I_BAThe foreground region only contains partial pixel points on the airplane target, namely an image I_BThe aircraft object in' is not completely segmented because of the region of interest ROI image I obtained in step 2)_AMiddle aircraft orderThere are marked cases where the amplitude of a few strong scattering points AH is much greater than the mean of the amplitudes of all scattering points I_AIs mean filtered in the filtered image I_B′(I_B) Forming a region with larger pixel amplitude on the aircraft target, namely a region with strong amplitude I on the aircraft target_B1′(I_B1) I.e. picture I_B′(I_B) Region of medium pixel amplitude greater than or equal to T1, I_B′(I_B) Removing I_B1′(I_B1) Region I outside_B2′(I_B2) Is less than T1, and I_B2′(I_B2) Is composed of two parts, one part is I_B′(I_B) On-target intensity-removal amplitude region I of aircraft in (1)_B1′(I_B1) Regions outside, called weak amplitude regions I on the aircraft target_B21′(I_B21) And the other part is I_B′(I_B) In a background area I other than aircraft targets_B22′(I_B22) I.e. I_B′＝I_B1′+I_B2′＝I_B1′+I_B21′+I_B22′，I_B＝I_B1+I_B2＝I_B1+I_B21+I_B22Albeit with I_B' Weak amplitude region I on aircraft target_B21' Pixel amplitude greater than I_B' background region I_B22' pixel amplitude, but due to I_B21' the pixel amplitude is still less than the threshold T1, so I_B' Weak amplitude region I on aircraft target_B21' and background region I_B22' is divided into binary images I_BABackground area of (1) only_B' Strong amplitude region I on aircraft target_B21' is divided into binary images I_BASo that the segmented binary image I_BAContains only the image I_B' the pixels in the region of strong amplitude on the aircraft target do not contain the pixels in the region of strong amplitude on the aircraft target, i.e. image I_B' the aircraft object is not completely segmented into binary images I_BAThe foreground region of (1);

(3c) for backup image I_BA' negation is performed and the result of the negation is used as a mask matrix for the blurred image I_BPerforming masking operation to obtain masked image

I_BThe non-zero region in (1) corresponds to I_BIn (1)_B2An area;

(3d) to I_BBPerforming threshold segmentation by using an OSTU method to obtain a segmentation threshold T2, and performing I pair by using T2_BBCarrying out binarization to obtain a binary image I_BCBinary image I obtained in this example_BCAs shown in FIG. 7, due to I_BBThe non-zero region in (1) corresponds to I_BIn (1)_B2Region, and I_BIn (1)_B2The region is composed of two parts, one part is I_BIn the region of weak amplitude I on the aircraft target_B21，I_B21Is greater than or equal to T2, and the other part is I_BIn a background area I other than aircraft targets_B22，I_B22Is less than T2, so I_BBIn the non-zero region of (1) corresponds to I_BIn the region of weak amplitude I on the aircraft target_B21Region is divided into binary images I_BCForeground region of (1), I_BBIs divided into a binary image I_BCBackground region of (1), i.e. binary image I_BCContains I_BIn the region of weak amplitude I on the aircraft target_B21The pixel point of (2);

subjecting the I obtained in step (3b)_B' thresholded binary image I_BAAnd I_BBThreshold-segmented binary image I_BCPerforming logical OR operation to obtain ROI image I_ABinary image I of_C(ii) a Essentially, a binary image I_BAForeground region and binary image I_BCAs a binary image I_CForeground region of (1), binary image I_BAContains an image I_B' Strong amplitude region I on aircraft target_B1' since I_B' is a_BA backup image of (2), then I_B′＝I_BThen, the binary image I_BAContains an image I_BIn the region of strong amplitude I on the aircraft target_B1Pixel point of (2), binary image I_BCContains I_BIn the region of weak amplitude I on the aircraft target_B21Pixel point of (2) then_B1And I_B21Is the union set of I_BAll the pixels on the aircraft target in (1), then the binary image I_CContains I_BThe method realizes the complete and effective extraction of the airplane target by all pixels on the airplane target, and the ROI image I obtained by the example_ABinary image I of_CAs shown in FIG. 8, it can be seen that the aircraft target is extracted more completely, and the binary image I of FIG. 8_CThe horizontal stripe noise is contained in the signal.

Step 4) obtaining a binary image I_CBinary image I after horizontal stripe noise suppression_D：

Horizontal stripe noise in binary image I_CThe middle is mainly represented as a set of 'bright spots' (pixel points with amplitude of 1) with a certain length, and usually, the horizontal stripe noise only appears near a target area. The main causes of the horizontal streak noise are the following, 1) the influence of side lobes of strong scattering points on the target; 2) caused by self-focusing errors; 3) there are rotating parts in the target. In a generic image, most of the cross-fringe noise is mainly caused by the side-lobe of the two-dimensional spread function of the strong scattering points on the object. The horizontal stripe noise seriously affects the quality of the image and seriously affects the extraction of the target length feature, so that the horizontal stripe noise suppression is required before the feature extraction. The specific steps of horizontal stripe noise suppression are as follows:

(4a) the initialization iteration number is T, the maximum iteration number is T, T is more than or equal to 1, in the example, T is 4, and the binary image I_CThe binary image after the t-th horizontal stripe noise suppression is I_D ^tAnd let t equal to 1, I_D ^t＝I_C；

(4b) Statistical binary image I_D ^tIn each row of pixel framesThe number of pixels with a value of 1 is obtained, and a pixel number sequence N is obtained as { N }₁,n₂,…,n_i,…,n_m′And judging whether the maximum value MAX and N' in N satisfy

Wherein n is_iRepresenting the number of pixels with the pixel amplitude value of 1 in the ith row, if so, representing a binary image I_D ^tIf there is horizontal stripe noise, executing step (4c), otherwise, representing the binary image I_D ^tIf there is no horizontal stripe noise, then I_D ^tAs for binary image I_C ^tBinary image I obtained by performing horizontal stripe noise suppression_DOutputting and ending the program;

(4c) carrying out first-order forward difference on the pixel number sequence N to obtain a first-order forward difference sequence delta N ═ delta N₁,Δn₂,…,Δn_d,…,Δn_m′-1Calculating the position d corresponding to the maximum value in the delta N₁，d₁Representing a binary image I_D ^tD in (d)₁The number of pixels having a pixel amplitude of 1 in the +1 row is increased most than that in the previous row, i.e., the d-th row₁+1 line corresponds to the binary image I_D ^tThe position d corresponding to the minimum value in Δ N is calculated as the start line of the horizontal streak noise in (1)₂，d₂Representing a binary image I_D ^tD in (d)₂The number of pixels having a pixel amplitude of 1 in the +1 row is reduced most than that in the previous row, i.e., the d-th row₂The lines corresponding to a binary image I_D ^tThereby determining a binary image I_D ^tLine range d where middle horizontal stripe noise is located₁+1～d₂And the width of the horizontal stripe noise, which is the number of lines occupied by the horizontal stripe noise, is W ═ d₂-d₁Wherein, Δ n_dRepresenting the difference, Δ n, between the number of pixels of adjacent rows having a pixel amplitude of 1_d＝n_d+1-n_d，0＜d₁≤m′-1，0＜d₂≤m′；

(4d) Selecting a binary image I_D ^tMiddle and largeBinary image I as small as l x n ″_CAAnd to I_CAPerforming convolution operation with all 1 convolution kernels h with the size of l multiplied by 1 to obtain a convolution result sequence V ═ V₁,V₂,…,V_j,…,V_n", wherein l ═ l₂-l₁+1，l₁And l₂Respectively representing binary images I_D ^tStarting and ending lines selected when selection is performed,/₁＝max(1,d₁+1-W)，l₂＝min(m′,d₂+ W), max (,) indicates the maximum value, min (,) indicates the minimum value, V_jIs represented by_CAThe convolution result of the jth column of (a) and h, j is more than 0 and less than or equal to n';

this example is for a binary image I_CBinary image I obtained by performing horizontal stripe noise suppression_DAs shown in fig. 9, it can be seen that the horizontal streak noise is effectively suppressed.

Step 5) obtaining a binary image I_DThe extraction result of the aircraft target length features in (1):

Carrying out zero equalization on the two-dimensional coordinate matrix G to obtain a zero-equalized two-dimensional coordinate matrix

Calculation of G₀Is G₀ ^TG₀Performing characteristic decomposition on S to obtain a characteristic value lambda of S₁And λ₂And λ₁＞λ₂Corresponding feature vector is v₁And v₂Then v is₁Direction vector of principal component, v₂As the direction vector of the sub-component, let the transformation matrix W be [ v [ ]₁ v₂]W is also called projection matrix, and is a two-dimensional coordinate matrix G which is equalized to zero by PCA (principal component analysis)₀Performing linear transformation by using the transformation matrix W to obtain a two-dimensional coordinate matrix after linear transformation

Wherein (x)_q,y_q) Representing a binary image I_DIs located at x_qLine y_qCoordinates of pixels of the column, (x)_q′,y_q') is (x)_q,y_q) Zero-averaged coordinates, (xs)_q,ys_q) Represents (x)_q′,y_q') coordinates after linear transformation, xs_qIs G₀Pixel (x) of (2)_q′,y_q') projection coordinates on the principal component, ys_qIs G₀Pixel (x) of (2)_q′,y_q') projection coordinates on the subcomponents, h denotes a binary image I_DTotal number of pixels of medium amplitude 1 [ ·]^TRepresenting a transpose operation;

(5b) according to G_SDetermining a line number q1 corresponding to the minimum value and a line number q2 corresponding to the maximum value of the 1 st column element in the binary image I_DHead coordinates (x) of middle aircraft target_q1,y_q1) And tail coordinates (x)_q2,y_q2) The length of a connecting line between the head and the tail of the airplane target is the length of the airplane target, and the distance dimension resolution ratio P in the step (1) is combined_rCalculating a binary image I_DLength of medium aircraft target:

the length feature extraction result of the airplane target obtained in this example is shown in fig. 10 (a), in which the head coordinate of the airplane target is (x)_q1,y_q1) Tail coordinate (x) 26,204_q2,y_q2)＝(195,27)，P_r0.3000m, the length of the aircraft target

The X coordinate in the image represents the column where the pixel is located, and corresponds to the second dimensional coordinate Y in the two-dimensional coordinate matrix G, and the Y coordinate in the image represents the row where the pixel is located, and corresponds to the first dimensional coordinate X in the two-dimensional coordinate matrix G.

The technical effects of the present invention will be further described with reference to simulation experiments.

1. Simulation conditions and contents:

the simulation is carried out on a Windows10 system with a CPU of Intel (R) core (TM) i7-10700F 2.90GHz and a memory of 32G, and the software platform of the used simulation experiment is MATLAB R2019 a. The data used in this experiment were: in the example, 1 ISAR image of the external field test aerial aircraft target is shown in fig. 2, the size of the amplitude image is 512 × 256, the distance dimension resolution is 0.3000m, the azimuth dimension resolution is 0.5390m, and 50 ISAR images of the external field test aerial aircraft target are 512 × 256, the distance dimension resolutions are 0.3000m, and the azimuth dimension resolutions are different.

The extraction results and the extraction accuracy of the method for extracting the target length features of the ISAR image are compared with those of the conventional method for extracting the target length features of the ISAR image, and the results are respectively shown in FIG. 10 and Table 1.

2. And (3) simulation result analysis:

TABLE 1

Referring to fig. 10, (a) in fig. 10 is a result of extracting an ISAR image aerial target length feature used in the present embodiment according to the present invention, and (b) in fig. 10 is a result of extracting an ISAR image aerial target length feature used in the present embodiment according to the prior artThe length feature extraction result of the aerial target of the ISAR image can be obtained from (a) in FIG. 10, the head coordinate of the aircraft target obtained by the invention is (26,204), the tail coordinate is (195,27), and the length is

As can be seen from fig. 10 (b), the aircraft object obtained in the prior art has a head coordinate of (26,211), a tail coordinate of (206,30), and a length of (d)

The real length of the airplane target in the ISAR image adopted in this embodiment is 73.7300m, and for the ISAR image adopted in this embodiment, the error of the length feature of the aerial target extracted by the present invention is 0.42%, and the error of the length feature of the aerial target extracted by the prior art is 3.87%.

Referring to table 1, table 1 shows the average extraction accuracy of the aircraft target length feature extraction results of 50 ISAR images, and as can be seen from table 1, the accuracy of the aircraft target length feature extraction result extracted by the method is improved by 4.63% compared with that of the existing method.

In conclusion, the extraction result of the aerial target length feature of the ISAR image of the actually measured data of the external field experiment proves that compared with the existing extraction method of the aerial target length feature of the ISAR image, the extraction result of the method is higher in precision, and the method has important practical significance.

The foregoing description is only an example of the present invention and should not be construed as limiting the invention in any way, and it will be apparent to those skilled in the art that various changes and modifications in form and detail may be made therein without departing from the principles and arrangements of the invention, but such changes and modifications are within the scope of the invention as defined by the appended claims.

Claims

1. An ISAR image aerial target length feature extraction method is characterized by comprising the following steps:

(1) preprocessing an ISAR image:

Calculating P_aAnd P_rRatio of

(3) obtaining ROI image I_ABinary image I of_C：

(3a) ROI image I of region of interest by Gaussian filtering method_ACarrying out smooth filtering to obtain a filtered image I_BAnd to I_BBackup is carried out to obtain I_BBackup image of I'_B；

(3b) To backup image I'_BPerforming threshold segmentation to obtain a threshold-segmented binary image I_BAAnd to I_BABackup is carried out to obtain I_BABackup image I of_BA′；

(4c) carrying out first-order forward difference on the pixel number sequence N to obtain a first-order forward difference sequence delta N ═ delta N₁，Δn₂，…Δn_d，…，Δn_m′-1And according to the position d corresponding to the maximum value in the delta N₁Position d corresponding to the minimum value₂Determining a binary image I_D ^tLine range d where middle horizontal stripe noise is located₁+1～d₂And the width W ═ d of the horizontal streak noise₂-d₁Wherein, Δ n_dRepresenting the difference, Δ n, between the number of pixels of adjacent rows having a pixel amplitude of 1_d＝n_d+1-n_d，0＜d₁≤m′-1,0＜d₂≤m′；

(4d) Selecting a binary image I_D ^tBinary image I with medium size of L multiplied by n ″_CAAnd count of I_CAThe number of pixels with the pixel amplitude of 1 in each row is obtained, and a pixel number sequence V ═ V is obtained₁，V₂，…V_j，…，V_n″Wherein L ═ L₂-L₁+1，L₁And L₂Respectively representing binary images I_D ^tThe starting and ending lines selected during selection, L₁＝max(1,d₁+1-W)，L₂＝min(m′,d₂+ W), max (,) indicates the maximum value, min (,) indicates the minimum value, V_jRepresenting a binary image I_CAThe number of pixels with the pixel amplitude value of 1 in the jth row of (1), j is more than 0 and less than or equal to n';

(4e) when V is_jWhen the image is less than or equal to W, the binary image is processed

D (d) of₁+1 line to d₂The amplitude value of the jth column pixel in the row is set to be 0, otherwise, the binary image is kept

D (d) of₁+1 line to d₂The amplitude of the jth column pixel in the row is unchanged, and the binary image after the t-th horizontal stripe noise suppression is obtained

2. the method for extracting the aerial target length feature of the ISAR image according to claim 1, wherein the ROI image with a size of m '× n' in the gray scale image I in step (1b) is extracted by:

(1b1) calculating the sum of the amplitudes of all pixels in each row in the gray level image I to obtain a row pixel amplitude sum sequence RS [ RS ]₁,rs₂,…,rs_x,…rs_m]And calculating the mean value rm of the elements in RS, wherein RS_xTo express a gray image I_xThe sum of all pixel amplitudes of a row;

(1b3) Calculating the sum of the amplitudes of all pixels in each column in the gray-scale image I to obtain a column pixel amplitude sum sequence CS [ CS ]₁,cs₂,…,cs_y,…cs_n]And calculating the mean value cm of elements in CS, wherein CS_yThe sum of all pixel amplitudes of the y-th column representing the gray-scale image I;

(1b4) computing each element CS in the CS_yDifference from cm gives the sequence C ═ C₁,c₂,…,c_y,…c_n]And the minimum value y of the positions corresponding to the elements larger than 0 in C_minAnd maximum value y_maxRespectively as the start column and the end column of the region of interest of the gray image I, wherein c_y＝cs_y-cm；

(1b5) Will be the xth of the gray image I_minGo to the x_maxLine, and y_minColumn to y_maxThe area image of the size m ' × n ' formed by the columns is determined as the region of interest ROI image of the gray image I, where m ' ═ x_max-x_min，n′＝y_max-y_min。

3. The method for extracting aerial target length features of ISAR images as claimed in claim 1, wherein in step (2), pixel interpolation is performed on the ROI image in the azimuth dimension, and pixel extraction is performed on the ROI image in the azimuth dimension, wherein the interpolation is performed by nearest neighbor interpolation, bilinear interpolation or bicubic interpolation, and the extraction is performed by nearest neighbor interpolation, bilinear interpolation or bicubic interpolation.

4. The method for extracting ISAR image aerial target length features as claimed in claim 1, wherein the step (3a) is performed by smoothing the ROI image of the region of interest by using a Gaussian filter method, the window size of the Gaussian filter is 5 x 5 to 9 x 9, and the standard deviation is 5 to 50.

5. The method for extracting aerial target length features of ISAR image according to claim 1, wherein the pair I in step (3b)_BBackup image I of_B' performing threshold segmentation, and the pair I described in step (3d)_BBAnd performing threshold segmentation by adopting an OSTU method.