KR101139588B1 - Method of discriminating class of target in synthetic aperture radar image and apparatus thereof - Google Patents

Abstract

The present invention relates to a method and apparatus for identifying a model of a target in a synthetic aperture radar image. According to the present invention, a method for identifying a model of a target in a composite aperture radar image includes: obtaining a composite aperture radar image for an unidentified target, normalizing the synthesized aperture radar image, and normalized synthesized aperture radar image Generating a polar image by processing the ultra-fine image, extracting the characteristic vector from the polar image, and comparing the extracted characteristic vector with the characteristic vector for the targets to be discriminated stored in the database. Identifying a step. As described above, according to the present invention, it is possible to discriminate not only the target but also the model of the target with high accuracy through various characteristic vectors. In addition, the feature vector according to an embodiment of the present invention may be combined with any classification algorithm (pattern recognition technique, neural network, fuzzy theory, etc.) and information fusion technique (information fusion). It can be applied directly to the field of Automatic Target Recognition (ATR) for identifying ground targets.

Description

FIELD OF THE INVENTION AND METHOD AND DEVICE FOR IDENTIFICING TYPE OF TARGET IN SYNTHETIC AROUND RADIOGRAPHIC TECHNICAL FIELD

The present invention relates to a method and apparatus for identifying a target type in a synthetic aperture radar image, and more particularly, to a synthetic aperture radar image capable of accurately distinguishing a ground target from a clutter in a synthetic aperture radar image. A method and apparatus for identifying a model of a target are provided.

In general, if you want to identify air, sea, and ground PIA targets, it is common that the target does not provide its own information. Therefore, it is essential to acquire the information on the target and identify what the target is based on. . Target identification requires the acquisition of unique physical characteristics of the target to be identified. Optical, infrared, sound, and radar signals are frequently used to achieve this characteristic.

Among them, radar uses light as a signal source, not light as an optical image. The use of radio waves has the advantage that the wavelength is longer than light and is not affected by clouds or sea fog. Therefore, the radar can provide information to enable the operation of all-weather operation because the information acquisition distance is longer than the optical and the weather is not limited. In particular, when used to identify ground targets, synthetic aperture radar (hereinafter referred to as "SAR image") is mainly used. SAR images obtained using radio waves are familiar to the human eye. The resolution is lower than that of an optical or electro-optical image that produces information.

FIG. 1 shows an optical image and a SAR image of a target on the same ground. The leftmost view of FIG. 1 is an optical image of a tank corresponding to the target, and the remaining four figures show the same tank at various angles. The captured SAR image is shown.

As shown in FIG. 1, the optical image may determine information about the target even from the human eye, that is, a tank, an aircraft, or a tank, but it is difficult to identify the type of the target by the SAR image.

In addition, SAR images can be targeted to detect because it shows an image of the area up to tens of km ² from km ^2, as well as natural clutter (natural clutter) and manmade clutter (man-made clutter) that appear in the video with do. Therefore, detection and identification of targets in SAR images are difficult to read by the reader's naked eye and require years of experience for accurate readings. In addition, there is a disadvantage in that real-time operation is impossible because it takes a lot of time to find, distinguish and identify targets included in an area of several tens of km ² . For this reason, computer signal processing using pattern recognition is essential to distinguish and identify targets from clutter in SAR images.

SUMMARY OF THE INVENTION An object of the present invention is to provide a method and apparatus for identifying a target type in a synthetic aperture radar image capable of distinguishing not only the target but also the class to which the target belongs with high accuracy.

According to an embodiment of the present invention, a method for identifying a model of a target in a synthetic aperture radar image may include obtaining a synthetic aperture radar image for an unidentified target, and performing the synthesis aperture radar image. Normalizing, generating a polar image by subjecting the normalized synthesized aperture radar image to extreme image processing, extracting a characteristic vector from the polar image, and discriminating the extracted characteristic vector stored in a database. Comparing with the characteristic vector for the targets, identifying the type of the unknown target.

In normalizing the composite aperture radar image, the composite aperture radar image is normalized using the following equation.

Where I _n (x _i , y _i ) is a normalized composite aperture radar image, I ₀ (x _i , y _i ) is a composite aperture radar image before normalization, and x and y are composite aperture radar images, respectively. The horizontal and vertical axis dimensions of the image are represented, and M and N represent the number of pixels for each of the x and y dimensions.

The generating of the polar image by performing the ultra-fine processing of the normalized synthetic aperture radar image may include generating a binary image by morphologically processing the normalized synthetic aperture radar image, detecting a center point from the binary image; And generating a polar image by sampling in a distance direction and an azimuth direction based on the center point of the normalized synthesized aperture radar image.

The feature vector extracted from the pole image may include a first feature vector for the distance mapping image of the pole image, a second feature vector for the azimuth mapping image of the pole image, and the pole image. And a third characteristic vector generated by compression), wherein the distance direction mapping image and the azimuth mapping image are each expressed by the following equation.

Here, I _r (r) is the mapping image in the direction of distance r, I _θ (θ) is the mapping image in the azimuth (θ) direction, I _p (r _m , θ _n ) represents the p-th pole image, R _min is the minimum sampling radius and R _max represents the maximum sampling radius.

The feature vector stored in the database includes a fourth feature vector for a distance mapping image of the pole images of the targets to be identified, a fifth feature vector for the azimuth mapping image of the pole images of the targets to be identified, and And a sixth characteristic vector generated by compressing the polar image of the target to be identified by image principal component analysis (IMPCA).

The identifying the unidentified target may include selecting a first group of targets having a high similarity by comparing the extracted first characteristic vector with the fourth characteristic vector, and aligning the polar images of the unidentified target to the second image. Extracting a feature vector and comparing the extracted second feature vector with the fifth feature vector of targets belonging to the first group to select targets of a second group with a high maximum normalized correlation coefficient, and the aligned The third characteristic vector is extracted from the polar image of the unidentified target, and the third characteristic vector is calculated with the sixth characteristic vector of targets belonging to the second group, so that a difference value among targets belonging to the second group is obtained. And selecting the smallest target as the model to which the unidentified target belongs.

The similarity is expressed by the following equation.

Where g ₁ (p _u , q _u ) is the distance mapping image for any angle of the unidentified target, and f ₁ ^T (p, q) is the fourth characteristic vector of the target to be identified stored in the database. It is an overall matrix, P is the number of classes of the target, and Q is the imaging angle of the target constituting the database.

An apparatus for identifying a model of a target in a composite aperture radar image according to another embodiment of the present invention includes a SAR image acquisition unit for obtaining a composite aperture radar image for an unidentified target, and a normalization for normalizing the synthesized aperture radar image. A polar image generating unit generating a polar image by performing extreme image processing on the normalized synthetic aperture radar image, a characteristic vector extracting unit extracting a characteristic vector from the polar image, and storing characteristic vectors of targets to be discriminated. And a discriminating unit for identifying the unidentified target by comparing the extracted characteristic vector with the characteristic vector for the target to be discriminated.

The normalizer may normalize the synthesized aperture radar image using the following equation.

Where I _n (x _i , y _i ) is a normalized composite aperture radar image, I ₀ (x _i , y _i ) is a composite aperture radar image before normalization, and x and y are composite aperture radar images, respectively. The horizontal and vertical axis dimensions of the image are represented, and M and N represent the number of pixels for each of the x and y dimensions.

As described above, according to the present invention, it is possible to discriminate not only the target but also the model of the target with high accuracy through various characteristic vectors. In addition, the feature vector according to the embodiment of the present invention can be combined with any classification algorithm (pattern recognition technique, neural network, fuzzy theory, etc.) and information fusion technique (information fusion), and radar, electro-optical sensor, infrared sensor, etc. It can be applied directly to the field of Automatic Target Recognition (ATR) for identifying ground targets.

1 illustrates an optical image and a SAR image of a target on the same ground.
2 is a diagram showing the configuration of an apparatus for identifying a model of a target in a SAR image according to an embodiment of the present invention.
3 is a flowchart illustrating a process of storing a characteristic vector of a target to be discriminated in a database according to an embodiment of the present invention.
4A shows a SAR image of a target to be identified according to an embodiment of the present invention.
FIG. 4B is a normalized image of the SAR image shown in FIG. 4A.
5 shows a polar lattice for the extreme projection process according to the embodiment of the present invention.
6 shows the geometry of the polar coordinate system correcting the polar lattice according to FIG. 5.
FIG. 7A shows a SAR image, and FIG. 7B shows an ISAR image.
FIG. 8 illustrates a binary image generated by morphologically processing the normalized SAR image illustrated in FIG. 4B.
FIG. 9A illustrates a polar grid positioned at the center of the target, and FIG. 9B illustrates a polar image generated by sampling FIG. 9A.
FIG. 10A illustrates a SAR image obtained by rotating the target shown in FIG. 9A by 90 °, and FIG. 10B illustrates a polar image generated by sampling FIG. 10A.
FIG. 11A shows the polar image and the extracted feature vector of the SAR image shown in FIG. 9B, and FIG. 11B shows the polar image and the extracted feature vector of the SAR image shown in FIG. 10B.
12 is a diagram showing the structure of a database for storing characteristic vectors according to an embodiment of the present invention.
13 is a flowchart illustrating a method of identifying a model of a target in a SAR image of a surveillance area according to an embodiment of the present invention.

DETAILED DESCRIPTION Embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention.

2 is a diagram showing the configuration of an apparatus for identifying a model of a target in a SAR image according to an embodiment of the present invention. The SAR image discrimination apparatus 200 according to FIG. 2 includes a SAR image acquisition unit 210, a normalization unit 220, a polar image generation unit 230, a feature vector extractor 240, a database 250, and a discrimination unit ( 260).

The SAR image acquisition unit 210 acquires a SAR image for the unidentified target. The normalizer 220 normalizes the acquired SAR image.

The pole image generator 230 generates a pole image by performing extreme mirror processing on the normalized SAR image. The feature vector extractor 240 extracts the feature vector from the polar image.

The database 250 stores characteristic vectors for targets to be discriminated. That is, the database 250 stores targets of various classes to be discriminated and characteristic vectors of SAR images obtained by photographing the targets from various angles.

The discriminating unit 260 identifies the model of the unidentified target by comparing the extracted characteristic vector with the characteristic vector for the target to be discriminated.

On the other hand, in order to directly use SAR images, SAR images for each target must be stored in a database. In this case, the same model is considered considering that SAR images are changed by depression angle and azimuth angle. Numerous SAR images are needed for the target. Therefore, not only a large amount of storage space of the database is required, but also a large amount of calculation time is required due to the large number of comparison images.

However, by extracting only the features of each target by applying a pattern recognition technique as in the embodiment of the present invention and storing them in the database 250, the storage space of the database can be reduced as well as the calculation time. In FIG. 3, a method of extracting a characteristic vector for a target to be discriminated and storing it in the database 250 will be described.

3 is a flowchart illustrating a process of storing a characteristic vector of a target to be discriminated in a database according to an embodiment of the present invention.

First, the SAR image acquisition unit 210 acquires a SAR image of a target to be discriminated for building a database (S310). Here, the target means a ground military target, and includes fixed targets such as various military bases as well as moving targets such as tanks, armored vehicles, missiles, and the like. SAR images are obtained by capturing images as the aircraft revolves around the ground at the point where one or more targets are located. The image acquired by the SAR image acquisition unit 210 is used to generate a feature vector constituting a database.

Next, the normalization unit 220 performs a normalization process of dividing the SAR image into total energy as a preprocessing process (S320). The radar image changes the signal level according to the distance between the sensor and the target, so the normalization process reduces this phenomenon. 4A illustrates a SAR image of a target to be identified according to an embodiment of the present invention, and FIG. 4B is an image obtained by normalizing the SAR image shown in FIG. 4A. As shown in FIGS. 4A and 4B, not only the target but also the shadow of the target (parts shown in black in FIGS. 4A and 4B) are also included in the SAR image.

If the SAR image is normalized as shown in FIG. 4B, the total energy of the normalized SAR image I _n (x _i , y _i ) becomes 1. Equation 1 shows a normalized SAR image (I _n (x _i , y _i )).

Here, a two-dimensional SAR image is obtained by (I ₀ (x _i , y _i )), i = 1,2,... , M, j = 1,2,... (X, y) represents the horizontal and vertical axis dimensions of the SAR image, and M and N represent the number of pixels for each of the x and y dimensions.

Next, the polar image generator 230 generates a polar image through a polar projection process in order to extract a feature vector used for target identification. Hereinafter, a method of generating a polar image from a SAR image will be described with reference to FIGS. 5 through 7B.

FIG. 5 shows a polar lattice for a polar projection process according to an embodiment of the present invention, and FIG. 6 shows the geometry of a polar coordinate system correcting the polar lattice according to FIG. 5.

First, scattering points far from the center of the SAR image are mainly due to diffraction at the edge of the target, which provides important information for target identification. Therefore, there is a need to minimize the loss of information for scattering points far from the center by keeping the sampling interval constant. The polar projection process obtains a polar image (I _p (r _m , θ _n )) by sampling in the direction of distance (r) and azimuth (θ) with a polar grid as shown in FIG. 5 at the center of the SAR image. Indicate the process. Where m = 1,2,... , N _r , n = 1,2,... , N _θ , N _r is the distance direction, and N _θ is the number of sampling in the azimuth direction. FIG. 6 illustrates the geometry of the polar image shown in FIG. 5, and shows a relationship as shown in Equation 2 between the variables of FIGS. 5 and 6.

Here, (x _c , y _c ) represents the coordinates of the center point of the SAR image, and Δr and Δθ represent sampling intervals in the distance direction and the azimuth direction, respectively. R _min is the minimum sampling radius and is selected to be greater than or equal to 0. R _max is the maximum sampling radius and is selected to include the target and the shadow of the target. Since the size of the target and the shadow of the target changes in the SAR image according to the change of the target size and the angle of inclination, the R _max is determined by considering the size of the shadow which is changed by the angle of inclination at the maximum length among the targets to be identified.

On the other hand, the difference between the extreme projection process between the SAR image and the ISAR image will be described with reference to FIGS. 7A and 7B. FIG. 7A shows a SAR image, and FIG. 7B shows an ISAR image.

In the ISAR image illustrated in FIG. 7B, the center point of the image may be selected as R _min = 0, similar to the SAR image. However, unlike SAR images, shadows and clutters are rarely seen in the SARR image, so the R _max of the ISAR microscopic method is selected to include only the target, but according to the SAR microscopic method, a value including the target and the shadow of the target is selected. R _max must be selected. This is because ISAR images are images of flying targets, but SAR images are images of ground targets.

In addition, R _max of ISAR pole mapping method is to ignore each R _max is the target size information of each image dalrajyeoseo the R _max because the selected value of the image containing only the target in the respective images. However, R _max of SAR microscopic projection uses target size information because it selects a constant value regardless of image.

5 to 7B, the polar image generating unit 230 detects the center of the normalized SAR image through a morphology process in order to convert the normalized SAR image into an ultrafine image (S330).

FIG. 8 illustrates a binary image generated by morphologically processing the normalized SAR image illustrated in FIG. 4B. In the ISAR image, since the pixels of a certain value or more are regarded as the target pixels, the center of the pixels can be easily obtained. In the case of the SAR image, the difference between the pixel value of the surrounding clutter and the pixel value of the target is not large. The process of finding the pixels is necessary.

A morphology process is performed on the normalized SAR image illustrated in FIG. 4B to generate a binary image in which the pixel of the target is 1, the shadow of the target and the pixel of the clutter are 0. Here, the center (x _c , y _c ) of the normalized SAR image is obtained by using the centers of the pixels of the target in the binary image as shown in Equation 3 below.

Here, (x _t , y _t ) is the coordinate value of the pixel of the target in the binary image.

When the center (x _c , y _c ) of the SAR image is determined, the polar image generator 230 generates a polar image through a polar projection process (S340). .

FIG. 9A illustrates a polar grid positioned at the center of the target, and FIG. 9B illustrates a polar image generated by sampling FIG. 9A. As shown in FIG. 9A, the polar grid illustrated in FIG. 5 may be positioned and sampled at the center of the normalized SAR image to obtain the polar image as illustrated in FIG. 9B. The polar image of Fig. 9b is N _r = N _θ = 100, R _max = 50 to create a polar image.

FIG. 10A illustrates a SAR image obtained by rotating the target shown in FIG. 9A by 90 °, and FIG. 10B illustrates a polar image generated by sampling FIG. 10A.

In the SAR image, the shadow appears behind the target, so when the radar is shown on the map when viewed from south to north and from west to east, the direction of the shadow behind the target is as shown in FIGS. 9A and 10A. Will be different. That is, the image may be rotated according to the SAR image generation method and the absolute position on the map of the radar and the target. However, when comparing FIG. 9B and FIG. 10B, it can be seen that the polar image shown in FIG. 10B is the same in the distance direction N _r as shown in FIG. 9B and merely shows a shape shifted in the azimuth direction N _θ . Can be.

Using the characteristic of the SAR image, the characteristic vector extractor 240 extracts the characteristic vector from the extreme image of the SAR image (S350).

The feature vectors extracted from the polar images are the distance projection image (I _r (r)), the azimuthal mapping image (I _θ (θ)), and image principal component analysis (IMPCA). It consists of the extreme image compressed.

FIG. 11A shows the polar image and the extracted feature vector of the SAR image shown in FIG. 9B, and FIG. 11B shows the polar image and the extracted feature vector of the SAR image shown in FIG. 10B.

When all the pixels of the polar image having the dimension of N _r × N _θ are used as the characteristic vectors, the capacity of the database increases. In order to reduce the dimension of the characteristic vector, the dimension of the polar image may be reduced by projecting the polar image in the distance direction N _r and the azimuth direction N _θ as shown in FIGS. 11A and 11B.

The distance mapping image I _r (r) and the azimuthal mapping image I _θ (θ) for the polar image generated by orthogonal SAR SAR images may be represented by Equations 4 and 5, respectively.

Here, I _r (r) is the mapping image in the direction of distance r, I _θ (θ) is the mapping image in the azimuth (θ) direction, I _p (r _m , θ _n ) represents the p-th pole image, R _min is the minimum sampling radius and R _max represents the maximum sampling radius.

As illustrated in FIGS. 11A and 11B, images projected in the distance direction N _r of two polar images have almost the same shape, and images projected in the azimuth direction N _θ are simply moved.

Since I _r (r) and I _θ (θ) alone are difficult to represent all the information of the polar image I _p (r _m , θ _n ), ISAR polar projection uses the principal component analysis (PCA) The feature vector representing the extreme image is extracted by minimizing the information loss of the polar image with a dimension smaller than the dimension of.

In general, since two-dimensional images have a high dimension, it is not easy to apply PCA to an image having a high number of dimensions. The reason for this is to generate a covariance matrix for the image, perform an eigenvalue decompositon process, and use a transform matrix (“B”) using an eigenvector corresponding to a eigenvalue having a large value. This is because it generates a compressed feature vector using For example, N _r × N if the image having the _θ dimension by applying a principal component analysis (PCA) assumed to generate the feature vector, N _r × N _θ D image size of the covariance matrix of the of the (N _r × N _{θ ) × (N r × N θ} ) have a dimension, and so it is difficult to decompose the eigenvalues of the covariance matrix.

Therefore, the image principal component analysis method (IMPCA) is applied to the SAR ultra-thin image method according to the embodiment of the present invention. First, an image covariance matrix (G) is represented by Equation 6 below.

는 식별하고자 하는 표적에 대한 극영상의 평균영상으로 m×n 차원이다. G를 고유값 분해(eigenvalue decompositon)를 하면 다음의 수학식 7과 같이 된다. Where G is the image covariance matrix in the n × n dimension, and A _j is the pole image of the j-th target of the m × n dimension.

Is the mean image of the pole image of the target to be identified, which is m × n dimension. When eigenvalue decompositon is performed on G, Equation 7 is obtained.

Here, A is a diagonal matrix having an eigenvalue of G as an element, and X is an n × n square matrix having an eigenvector of G as an element.

By transforming the eigenvalues of the image covariance matrix, a transformation matrix B may be generated as shown in Equation 8 by using an eigenvector having d (d <n) eigenvalues.

That is, the capacity of the m × n-dimensional polar image can be compressed into the m × d-dimensional polar image through the IMPCA.

In addition, a matrix having l (= m * d) elements is generated by multiplying the transformation matrix B by the polar image (A, m × n) and then vectorizing the characteristic vector (I _IMPCA (l)). ) Can be created as a column vector.

ISAR images show only the target in the region of interest with little clutter. However, because SAR images are ground-based, there are many clutters on the surface, plants, trees, and artificial structures. As a result, when the eigenvalues of the covariance matrix (G) of the ISAR image are calculated, most of the energy is present in a few eigenvalues, and the eigenvalues having the values represent the values of the target pixels. However, when calculating the eigenvalues for the covariance matrix of SAR images, there are many eigenvalues in which most of the energy is not only large eigenvalues but also small values. Because it contains information about clutter images, you should select the appropriate number of eigenvalues.

Here, N _r , N _θ , and l values are determined by the resolution of the SAR image and the area occupied by the target in the SAR image.

As a result, the database 250 compresses the feature vector of the distance mapping image I _r (r) for the extreme image, the feature vector of the azimuth mapping image I _θ (θ), and the image principal component analysis method (IMPCA). _{Stores the} characteristic vector I _IMPCA (l).

That is, the characteristic vector for the q-th angle for the p-th target is composed of f ₁ (p, q), f ₂ (p, q), f ₃ (p, q), and I _r (r) and I _θ , respectively. (θ) and the characteristic vector of I _IMPCA (l). In the embodiment of the present invention, N _r It is assumed that a feature vector having 200 elements is selected to represent one SAR image by selecting = N _θ = 50 and l = 100. This is only 200/120 ² = about 1.4% compared to using 120 × 120 SAR images directly, which can significantly reduce the storage capacity of the database.

The feature vector generated in this manner is used to discriminate the unknown SAR image, and the generated feature vector is stored in the database 250 (S360).

12 is a diagram showing the structure of a database for storing characteristic vectors according to an embodiment of the present invention. As shown in FIG. 12, three characteristic vectors according to various angles are stored for each target. In the case of the p-th target, the three characteristic vectors f ₁ (p, q) and f ₂ (p, q) according to the q-th angle are stored. , f ₃ (p, q) is stored in the database 250.

Hereinafter, a method of discriminating a target from a SAR image photographed during actual reconnaissance will be described using the characteristic vector of the target stored in the database 250.

13 is a flowchart illustrating a method of identifying a model of a target in a SAR image of a surveillance area according to an embodiment of the present invention.

First, when the reconnaissance plane captures a SAR image of an unidentified target while scouting over a surveillance area, the SAR image acquisition unit 210 receives a SAR image including an unidentified target (S1310). If the SAR image discrimination apparatus 200 is included in the reconnaissance plane, the SAR image acquisition unit 210 may directly acquire and acquire a SAR image, and the SAR image discrimination apparatus 200 is not included in the reconnaissance aircraft. Otherwise, the SAR image acquisition unit 210 may receive and acquire a SAR image photographed from the reconnaissance plane.

Next, the normalization unit 220 performs a normalization process of dividing the SAR image into total energy as a preprocessing process (S1320). Since the preprocessing process of the normalization unit 220 has been described with reference to FIGS. 4A and 4B, redundant description thereof will be omitted.

The pole image generator 230 detects the center of the normalized SAR image through a morphology process in order to convert the normalized SAR image into an ultrafine image (S1330). When the center of the SAR image is determined, the pole image generator 230 generates a pole image through a process of extreme projection (S1340). .

Since the polar image generating unit 230 generates the polar image from the normalized SAR image through the extreme mirror process, the overlapping description will be omitted.

Next, the feature vector extractor 240 extracts the feature vectors f ₁ (p, q) and f ₂ (p, q) from the polar image of the SAR image (S1350). That is, as described with reference to FIGS. 11A and 11B, by projecting the polar image in the distance direction N _r and the azimuth direction N _θ , the characteristic vector f ₁ for the distance direction mapping image I _r (r) is projected. The feature vectors f ₂ (p, q) for (p, q) and the azimuth mapped image I _θ (θ) are extracted.

The discriminating unit 260 discriminates the target primarily by comparing the extracted characteristic vector f ₁ (p, q) with the characteristic vector stored in the database 250 (S1360). That is, the discriminating unit 260 selects only targets having high similarity by calculating the similarity (C (p, q)) of the distance mapping image I _r (r) between the database and the target as shown in Equation 10.

Where g ₁ (p _u , q _u ) is the distance (N _r ) mapping image for any angle of the unidentified target, and f ₁ ^T (p, q) is the characteristic of the target to be identified stored in the database. Is the entire matrix of the vector f ₁ (p, q). P represents the number of classes of the target, and Q represents the imaging angle of the target constituting the database.

In this way, the discriminating unit 260 calculates C (p, q) and sorts in descending order, selecting only the target corresponding to the highest η% as an index of the database and setting it as an S1 group.

Next, the discriminating unit 260 secondaryly discriminates the target by comparing the extracted characteristic vector f ₂ (p, q) with the characteristic vector stored in the database 250 (S1370).

Here, a discrete correlation theorem is used to align the azimuth mapping image I _θ (θ) in the azimuth direction (N _θ ). That is, as described with reference to FIGS. 11A and 11B, since the image projected in the azimuth direction N _θ of two polar images captured from different angles is simply moved, the azimuth mapped image I _θ ( θ). Computation time can be reduced by applying discrete correlation theory to the S1 group that has undergone the primary discrimination process (S1360).

Equations 11 to 16 are for explaining a process of aligning the azimuthal sagittal image I _θ (θ). If an arbitrary vector a having k elements is defined, cyclic shift S ^h can be expressed as in Equations 11 and 12 below.

Then, using Equation 12, a difference between an unidentified target and a target to be identified as shown in Equation 13 is obtained.

g ₂ (p _u , q _u ) is the azimuth (N _θ ) mapping image for any angle of the unidentified target, and f ₂ (p, q) is the characteristic vector of the target to be identified stored in the database. When MS (h) shown in Equation 13 is minimum, the similarity between the two vectors is maximum.

The right end of Equation 13 may be expressed as Equation 14 below.

Where | f ₂ (p, q) | ² and | g ₂ (p, q) | ^{Since 2} is a fixed value, when the rightmost component in Equation 14 becomes the maximum, Equation 13 becomes the minimum. Therefore, some of the rightmost components of Equation 14 are expressed as Equation 15.

In order to minimize the MS (h) shown in Equation 13, Equation 15 should be maximized, so h is calculated to be the maximum of λ (h) in Equation 15.

Performing FFT (fast Fourier transform) on Equation 15 is as follows.

Where F ₂ (p, q) = FFT (f ₂ (p, q)), G ₂ (p, q) = FFT (g ₂ (p, q)), and G ₂ ^* (p, q) is G _Is a conjugate complex of ₂ (p, q). In Equation 16, h is determined to be the maximum of the lambda value.

Since the rotation angle of the azimuth direction (N _θ ) in the SAR image is represented by h in the polar image, the azimuthal sagittal image I _θ (θ) can be aligned without performing a circular movement process, and the discriminating unit 260 is mathematical. As shown in Equation 17, only the targets having a high correlation coefficient value are selected by calculating the maximum normalized correlation coefficient M (p ', q') of the distance mapping image I _θ (θ) between the database and the target.

p 'represents a specific target selected from the S1 group and q' represents an imaging angle of the specific target selected from the S1 group.

That is, the discriminating unit 260 calculates M (p ', q') and sorts in descending order as shown in Equation 17, selecting only the targets of the S1 group belonging to the upper γ% among the entire S1 groups, and selecting S2 as an index of the database. Set as a group.

In the primary identification process (S1360) and the secondary identification process (S1370), the η and γ values can be determined somewhat flexibly for SAR identification. The reason is that there are shadows and clutter of the target in the SAR image, and the clutter changes in the distance mapping image and the azimuthal mapping image according to the plain, forest, and city.

Next, the discriminating unit 260 discriminates the target on a third basis by comparing the extracted feature vector f ₃ (p, q) with the feature vector stored in the database 250 (S1380).

The polar image of the unidentified target is aligned with the SAR polar image of the target to be identified belonging to the S2 group by using the rotation angle h stored in the second identification process (S1360). That is, if the pole image of the unidentified target is moved by the rotation angle h and then vectorized, the compressed feature vector I _IMPCA (l) may be generated by the image principal component analysis method (IMPCA).

Then, each of g ₃ (p _u , q _u ) for the unidentified target is calculated using Equation 18 using the transformation matrix B used to obtain f ₃ (p, q).

p "represents the specific target selected from the S2 group and q" represents the imaging angle of the specific target selected from the S2 group.

For convenience of description, in Equation 18, g ₃ (p _u , q _u ) is a feature value of the extreme optical image compressed by the IMPCA of an unidentified target, and f ₃ (p ", q") is expressed in a database. It is defined as the characteristic vector value compressed by the extreme principal component analysis (IMPCA) of the target to be stored.

When (p ", q") is included in the S2 group, the class of the unidentified target is selected as p ₀ if d (p ₀ , q ₀ ) ≤d (p ", q"). That is, p ₀ is selected as the same target as the unidentified target so that d (p ", q") is minimized.

As described above, according to the present invention, it is possible to discriminate not only the target but also the model of the target with high accuracy through various characteristic vectors. In addition, the feature vector according to an embodiment of the present invention may be combined with any classification algorithm (pattern recognition technique, neural network, fuzzy theory, etc.) and information fusion technique (information fusion). It can be applied directly to the field of Automatic Target Recognition (ATR) for identifying ground targets.

Although the preferred embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concepts of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

200: SAR image discrimination apparatus 210: SAR image acquisition unit
220: normalization unit 230: polar image generation unit
240: feature vector extractor 250: database
260: discriminating part

Claims (13)

Obtaining a synthetic aperture radar image for the unidentified target,
Normalizing the composite aperture radar image;
Extreme image processing of the normalized synthetic aperture radar image to generate a polar image;
Extracting a feature vector from the polar image, and
Comparing the extracted characteristic vector with characteristic vectors of targets to be discriminated stored in a database, and identifying a model of the unidentified target,
The extreme image processing of the normalized synthetic aperture radar image may be performed to generate a polar image.
Generating a binary image by morphologically processing the normalized synthetic aperture radar image,
Detecting a center point from the binary image, and
And generating a pole image by sampling in a distance direction and an azimuth direction based on the center point of the normalized synthetic aperture radar image. The method of claim 1,
Normalizing the composite aperture radar image may include:
A method of identifying a target model in a composite aperture radar image for normalizing the composite aperture radar image using the following equation:

Where I _n (x _i , y _i ) is a normalized composite aperture radar image, I ₀ (x _i , y _i ) is a composite aperture radar image before normalization, and x and y are composite aperture radar images, respectively. The horizontal and vertical axis dimensions of the image are represented, and M and N represent the number of pixels for each of the x and y dimensions. delete The method of claim 2,
The feature vector extracted from the polar image is
A first characteristic vector of the distance mapping image of the polar image;
A second characteristic vector of the azimuth mapping image of the polar image, and
A third characteristic vector generated by compressing the extreme image by an image principal component analysis method (IMPCA);
The distance mapping image and the azimuth mapping image, respectively, a method of identifying a model of a target in a composite aperture radar image represented by the following equation:

Here, I _r (r) is the mapping image in the direction of distance r, I _θ (θ) is the mapping image in the azimuth (θ) direction, I _p (r _m , θ _n ) represents the p-th pole image, R _min is the minimum sampling radius and R _max represents the maximum sampling radius. The method of claim 4, wherein
The characteristic vector stored in the database is
A fourth characteristic vector of the distance mapping image of the pole image of the targets to be identified;
A fifth characteristic vector of the azimuth mapping image of the polar images of the targets to be identified; and
And a method of identifying a target type in a synthetic aperture radar image including a sixth characteristic vector generated by compressing the polar image of the targets to be identified by image principal component analysis (IMPCA). The method of claim 5,
Identifying the unidentified target,
Selecting the first group of targets having a relatively high similarity by comparing the extracted first characteristic vector with the fourth characteristic vector;
Extracting the second characteristic vector by aligning the polar images of the unidentified target, and comparing the extracted second characteristic vector with the fifth characteristic vector of the targets belonging to the first group, the second having the highest normalized correlation coefficient; Selecting targets in the group, and
Extracting the third characteristic vector from the polar images of the unidentified targets aligned, and calculating the third characteristic vector with the sixth characteristic vector of targets belonging to the second group, thereby selecting among the targets belonging to the second group And selecting the target having the smallest difference value as the model to which the unidentified target belongs. The method of claim 6,
The similarity is to identify the model of the target in the synthetic aperture radar image expressed as the following equation:

Where g ₁ (p _u , q _u ) is the distance mapping image for any angle of the unidentified target, and f ₁ ^T (p, q) is the fourth characteristic vector of the target to be identified stored in the database. It is an overall matrix, P is the number of classes of the target, and Q is the imaging angle of the target constituting the database. A SAR image acquisition unit for obtaining a synthetic aperture radar image for an unidentified target,
A normalizer for normalizing the synthesized aperture radar image,
Polar image generating unit for generating a polar image by the extreme image processing of the normalized synthetic aperture radar image,
Feature vector extraction unit for extracting a feature vector from the polar image,
A database that stores characteristic vectors for the targets to discriminate, and
Comprising a discriminating unit for identifying the unidentified target by comparing the extracted characteristic vector with the characteristic vector for the target to be discriminated,
The polar image generating unit,
The binary image is generated by morphologically processing the normalized composite aperture radar image, the center point is detected from the binary image, and then sampled in a distance direction and an azimuth direction based on the center point of the normalized synthetic aperture radar image. An apparatus for identifying a model of a target in a synthetic aperture radar image for generating an image. The method of claim 8,
The normalization unit,
An apparatus for identifying a model of a target in a composite aperture radar image for normalizing the composite aperture radar image using the following equation:

Where I _n (x _i , y _i ) is a normalized composite aperture radar image, I ₀ (x _i , y _i ) is a composite aperture radar image before normalization, and x and y are composite aperture radar images, respectively. The horizontal and vertical axis dimensions of the image are represented, and M and N represent the number of pixels for each of the x and y dimensions. delete 10. The method of claim 9,
The feature vector extracted from the polar image is
A first characteristic vector of the distance mapping image of the polar image;
A second characteristic vector of the azimuth mapping image of the polar image, and
A third characteristic vector generated by compressing the extreme image by an image principal component analysis method (IMPCA);
The apparatus for identifying a model of a target in the composite aperture radar image represented by the distance mapping image and the azimuth mapping image, respectively:

Here, I _r (r) is the mapping image in the direction of distance r, I _θ (θ) is the mapping image in the azimuth (θ) direction, I _p (r _m , θ _n ) represents the p-th pole image, R _min is the minimum sampling radius and R _max represents the maximum sampling radius. The method of claim 11,
The characteristic vector stored in the database is
A fourth characteristic vector of the distance mapping image of the pole image of the targets to be identified;
A fifth characteristic vector of the azimuth mapping image of the polar images of the targets to be identified; and
An apparatus for identifying a model of a target in a synthetic aperture radar image comprising a sixth characteristic vector generated by compressing the polar image of the target to be identified by an image principal component analysis (IMPCA). The method of claim 12,
The discriminating unit,
Comparing the extracted first characteristic vector with the fourth characteristic vector to select targets of a first group having a relatively high similarity,
Extracting the second characteristic vector by aligning the polar images of the unidentified target, and comparing the extracted second characteristic vector with the fifth characteristic vector of the targets belonging to the first group, the second having the highest normalized correlation coefficient; Select targets in the group,
Extracting the third characteristic vector from the polar images of the unidentified targets aligned, and calculating the third characteristic vector with the sixth characteristic vector of targets belonging to the second group, thereby selecting among the targets belonging to the second group A device for identifying a model of a target in a synthetic aperture radar image for selecting a target having a smallest difference value as a model to which the unknown target belongs.

Images