JP2015021824A - Classification and identification filter generator and target classification and identification device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate a classification and identification filter with high classification and identification performance of a target without increasing a calculation amount.SOLUTION: A classification and identification filter generator is provided with a teacher image positioning unit 3 which calculates adaptation among a plurality of teacher images by using a spectral signal of a frequency region transformed by a two-dimensional Fourier transformation unit 2, classifies the plurality of teacher images into one or more clusters on the basis of the adaptation, and positions the teacher images belonging to the same cluster, in which a classification and identification filter synthesis unit 4 synthesizes the teacher images positioned by the teacher image positioning unit 3 with one another for every cluster classified by the teacher image positioning unit 3, and generates a classification and identification filter.

Description

  The present invention relates to a target class identification device that compares a radar observation image of an unknown target with a class identification filter to determine a target type, and a class identification filter generation apparatus that generates the class identification filter. .

A target class identification method that determines the type of target is known by performing pattern matching between the observation image of a static target obtained by radar observation and a teacher image of a plurality of target targets that are obtained in advance. It has been.
As the teacher image, an image calculated in advance by computer simulation, an image cut out from a radar observation image of a known target that has been observed in advance (an image of a portion in which the target is reflected), or the like is used.

When an image cut out from a radar observation image of a target of a known model is used as a teacher image, even if the radar observation image is an image obtained by observing the same target, the image differs greatly if the direction of the target is different. Therefore, for each target, it is necessary to prepare images observed from all angles as teacher images in advance. Therefore, there is a problem that a large memory for storing a large number of teacher images is required.
In addition, there is a problem that the amount of calculation increases because it is necessary to perform pattern matching with a large number of teacher images when determining the type of target.

As a method for solving these problems, Non-Patent Document 1 below includes a class identification filter by integrating a plurality of observation images for the same target observed in a plurality of observation specifications (target main axis direction (orientation)). A method of generating and performing pattern matching with the class identification filter is disclosed.
If the method disclosed in the following Non-Patent Document 1 is used, it is not necessary to prepare images observed from all angles in advance as teacher images for each target. There is no need to perform matching.
For this reason, the capacity of the memory for storing the teacher image can be reduced, and the speed of the class identification process can be increased.

However, in order to generate a class identification filter, before integrating a plurality of observation images, preprocessing (for example, luminance adjustment processing, size normalization processing, center position alignment processing) is performed on each observation image. Etc.). In particular, the center position alignment process generally needs to be performed manually.
When an image generated by computer simulation is used as the teacher image, since the target position in the image is known, it is easy to align the center position.
However, when using an image cut out from a radar observation image observed in advance as the teacher image, the position of the target in the image is unknown. It must be aligned, and work efficiency will deteriorate.

Here, FIG. 7 is a block diagram showing a class identification filter generating apparatus to which the method disclosed in Non-Patent Document 1 is applied.
In FIG. 7, a teacher image data file 101 stores teacher image data indicating a target teacher image cut out from a plurality of target observation images observed under different observation conditions.
The two-dimensional Fourier transform unit 102 performs a process of performing a two-dimensional Fourier transform on each of the plurality of teacher image data stored in the teacher image data file 101.
The class identification filter synthesizing unit 103 generates a class identification filter 104 called a MINACE (Minimum Noise Correlation Energy) filter by synthesizing two-dimensional Fourier transform results of a plurality of teacher image data by the two-dimensional Fourier transform unit 102. To implement.

FIG. 8 is a conceptual diagram showing a class identification filter generation process by the class identification filter synthesis means 103.
Hereinafter, processing contents of the class identification filter generation device of FIG. 7 will be described.
When two-dimensional Fourier transform means 102 receives teacher image data indicating N teacher images from the teacher image data file 101, the two-dimensional Fourier transform means 102 performs two-dimensional Fourier transform on the N teacher image data.
The two-dimensional Fourier transform means 102 performs two-dimensional Fourier transform on the N pieces of teacher image data, raster scans the two-dimensional spectrum signal, which is the result of the two-dimensional Fourier transform, into a vector, and vectorizes the two-dimensional spectrum signal x _1. , X ₂ ,..., X _N are output to the class identification filter synthesis means 103.
In FIG. 8, teacher image data indicating four teacher images A, B, C, and D are input to the two-dimensional Fourier transform means 102, and four vectorized two-dimensional spectrum signals x ₁ , x ₂ , x _{3, x N} are output to the class identification filter synthesis unit 103.

ただし、

Upon receiving the vectorized two-dimensional spectrum signals x ₁ , x ₂ ,..., X _N from the two-dimensional Fourier transform unit 102, the class identification filter synthesis unit 103 receives the two-dimensional spectrum signals x ₁ , x ₂ , .., X _N is substituted into the following equation (1) to generate a MINACE filter h that is a class identification filter 104.

However,

式（４）において、Ｗは雑音の共分散行列であり、全ての教師画像の電力スペクトルの値が雑音レベルの定数倍（ｃ倍）を下回る場合、教師画像の電力スペクトルの代わりに、雑音レベルの定数倍（ｃ倍）の値で置き換える。
なお、式（４）の最後のｃＷは、生成されるＭＩＮＡＣＥフィルタｈの雑音耐性を高めるために導入されている。 T is a diagonal matrix, and the diagonal component is given by the maximum value of the power spectrum of all teacher images, as shown in the following equation (4).

In Expression (4), W is a noise covariance matrix, and when the power spectrum values of all the teacher images are less than a constant multiple (c times) of the noise level, the noise level is used instead of the power spectrum of the teacher image. Replace with a value that is a constant multiple of (c times).
Note that the last cW in the equation (4) is introduced in order to increase noise resistance of the generated MINACE filter h.

Here, the MINACE filter h is given in the form of a “weighted average” of the vectorized two-dimensional spectrum signals x ₁ , x ₂ ,..., X _N and is defined in the two-dimensional spectrum space. Yes.
Therefore, when the MINACE filter h is subjected to two-dimensional inverse Fourier transform, the shape of the target class identification filter in the image region is obtained as shown in the class identification filter 104 of FIG.
Strictly speaking, since the MINACE filter h is given in the form of a vector, the two-dimensional spectrum signal x ₁ , It is necessary to perform a two-dimensional inverse Fourier transform after returning the MINACE filter h to a two-dimensional signal in the reverse manner of obtaining x ₂ ,..., x _N.

In the synthesizing process of the class identification filter (MINACE filter h) shown in FIG. 8, the alignment process of a plurality of teacher images is manually performed in advance.
However, in the manual alignment process, as illustrated in FIG. 8, the plurality of teacher images cannot be accurately aligned and may be combined while being shifted. This problem becomes prominent when the teacher image is an image cut out from a radar observation image observed in advance.
When the accuracy of the alignment is insufficient, the obtained class identification filter 104 has a shape with a wide spread as depicted in FIG. Will cause.

  As a method for improving the alignment accuracy of a plurality of teacher images, a reference teacher image is arbitrarily selected as a reference image from among a plurality of teacher images, and the remaining teacher images are shifted by one pixel from the reference image. Patent Document 1 below discloses a method for performing registration when the correlation coefficient between the reference image and the teacher image is maximized.

JP2012-127920A

Rohit Patnaik, “Distortion-invariant Kernel Correlation Filters for General Object Recognition”, A project report for the degree of Master of Science, Carnegie Mellon University, Department of Electrical and Computer Engineering, April 10,1995

Since the conventional class identification filter generation device is configured as described above, the alignment accuracy can be improved if alignment is performed where the correlation coefficient between the reference image and the teacher image is maximized, Since this alignment processing is processing in an image region, there is a problem that requires a large amount of calculation.
In addition, when the orientations of targets existing in a plurality of teacher images are different, alignment can be performed with high accuracy even when alignment is performed where the correlation coefficient between the reference image and the teacher image is maximized. There was a problem that could not be done.

Further, since all the images of all angles are synthesized, the class identification filter 104 generated by the class identification filter synthesis unit 103 becomes an over-generalized class identification filter. For this reason, in the case of observation images included in the same target class, the fitness level is high regardless of the angle, but the fitness level of other targets is also high. was there.
For this reason, it is necessary to determine in advance a target angle range to be combined into one class identification filter, but it is generally difficult to determine in advance an angle range that can be combined. In addition, since the radar observation image is observed like a collection of point images depending on the resolution, it is not always easy to determine the target direction from the image. In other words, since it is difficult to accurately know the target orientation included in the acquired teacher image, even if the angle range to be synthesized is determined in advance, it is accurately determined whether it is included in the angle range. It is difficult to confirm.

The present invention has been made to solve the above-described problems, and provides a class identification filter generation device capable of generating a class identification filter with high target class identification performance without increasing the amount of calculation. For the purpose.
In addition, the present invention provides a class identification filter generation device capable of generating a class identification filter having high target class identification performance even when the target orientations existing in a plurality of teacher images are different. Objective.
Another object of the present invention is to provide a target class identification device that can accurately identify the type of a target by using a class identification filter having a high target class identification performance.

  A class identification filter generation device according to the present invention is stored by a teacher image storage unit that stores a target teacher image cut out from a plurality of target observation images observed under different observation conditions, and a teacher image storage unit. Frequency domain conversion means for converting each of the teacher image data indicating a plurality of teacher images into frequency domain spectrum signals, and using the frequency domain spectrum signals converted by the frequency domain conversion means, between the plurality of teacher images There is provided a teacher image alignment means for calculating the fitness, classifying the plurality of teacher images into one or more clusters based on the fitness, and aligning the teacher images belonging to the same cluster. For each cluster classified by the teacher image alignment means, the identification filter synthesis means performs the rank by the teacher image alignment means. Combined is obtained so as to generate a synthesized and kind identifying filter teacher images with each other is performed.

  According to the present invention, the degree of adaptation between a plurality of teacher images is calculated using the frequency domain spectrum signal converted by the frequency domain conversion means, and the plurality of teacher images are converted into one or more teacher images based on the degree of adaptation. Teacher image alignment means for classifying the teacher images belonging to the same cluster and classifying the teacher images is provided for each cluster classified by the teacher image alignment means. Since the class identification filter is generated by combining the teacher images that have been aligned by the matching unit, a class identification filter with high target class identification performance is generated without increasing the amount of computation. There is an effect that can.

It is a block diagram which shows the target class identification device which mounts the class identification filter production | generation apparatus by Embodiment 1 of this invention. It is a conceptual diagram which shows the production | generation process of the class identification filter by the teacher image position alignment part 3 and the class identification filter synthetic | combination part 4. FIG. It is a block diagram which shows the teacher image position alignment part 3 of the class identification filter production | generation apparatus by Embodiment 1 of this invention. It is a block diagram which shows the target class identification apparatus which mounts the class identification filter production | generation apparatus by Embodiment 2 of this invention. It is a block diagram which shows the teacher image position alignment part 20 of the class identification filter production | generation apparatus by Embodiment 2 of this invention. It is a block diagram which shows the teacher image position alignment part 20 of the class identification filter production | generation apparatus by Embodiment 3 of this invention. It is a block diagram which shows the classification identification filter production | generation apparatus to which the method currently disclosed by the nonpatent literature 1 is applied. It is a conceptual diagram which shows the production | generation process of the class identification filter by the class identification filter synthetic | combination means 103. FIG.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a target class identification device equipped with a class identification filter generation device according to Embodiment 1 of the present invention.
In FIG. 1, the teacher image storage unit 1 includes a storage device such as a RAM or a hard disk, and a plurality of targets observed under different observation conditions (for example, observation parameters such as a target main axis direction (orientation)). The target teacher images cut out from the observed images are stored. The teacher image storage unit 1 constitutes teacher image storage means.
The two-dimensional Fourier transform unit 2 is composed of, for example, a semiconductor integrated circuit on which a CPU is mounted, a one-chip microcomputer, or the like, and applies to teacher image data indicating a plurality of teacher images stored by the teacher image storage unit 1. By performing a two-dimensional Fourier transform process, a plurality of teacher image data is converted into spectrum signals in the frequency domain. Note that the two-dimensional Fourier transform unit 2 constitutes frequency domain transforming means.

  The teacher image alignment unit 3 is composed of, for example, a semiconductor integrated circuit on which a CPU is mounted, a one-chip microcomputer, or the like, and a plurality of frequency domain spectral signals converted by the two-dimensional Fourier transform unit 2 are used. The degree of matching between the teacher images is calculated, the plurality of teacher images are classified into one or more clusters on the basis of the degree of matching, and the alignment processing of the teacher images belonging to the same cluster is performed. The teacher image alignment unit 3 constitutes a teacher image alignment unit.

The class identification filter synthesis unit 4 is constituted by, for example, a semiconductor integrated circuit on which a CPU is mounted, a one-chip microcomputer, or the like, and for each cluster classified by the teacher image registration unit 3, the teacher image registration unit 3 A process for generating a class identification filter by combining the teacher images that have been aligned by the above is performed. The class identification filter synthesis unit 4 constitutes a class identification filter synthesis unit.
The class identification filter storage unit 5 is composed of a storage device such as a RAM or a hard disk, and stores the class identification filter generated by the class identification filter synthesis unit 4.

  The target type determination unit 6 includes, for example, a semiconductor integrated circuit on which a CPU is mounted, a one-chip microcomputer, or the like, and is stored in the observed image of the target to be classified and the class identification filter storage unit 1 A process for determining the type of target is performed by collating with the above-described class identification filter. The target type determination unit 6 constitutes a target type determination unit.

In the example of FIG. 1, a teacher image storage unit 1, a two-dimensional Fourier transform unit 2, a teacher image alignment unit 3, a class identification filter synthesis unit 4, a class identification filter storage unit 5, and a target, which are components of the target class identification device. Although each of the type determination units 6 is assumed to be configured with dedicated hardware, the target class identification device may be configured with a computer.
When the target class identification device is configured by a computer, the teacher image storage unit 1 and the class identification filter storage unit 5 are configured on the internal memory or the external memory of the computer, and the two-dimensional Fourier transform unit 2, the teacher image alignment The program describing the processing contents of the unit 3, the class identification filter synthesis unit 4 and the target type determination unit 6 is stored in the memory of the computer, and the CPU of the computer executes the program stored in the memory. do it.
FIG. 2 is a conceptual diagram illustrating a class identification filter generation process performed by the teacher image registration unit 3 and the class identification filter synthesis unit 4.

FIG. 3 is a block diagram showing the teacher image alignment unit 3 of the class identification filter generation device according to Embodiment 1 of the present invention.
In FIG. 3, the complex conjugate multiplier 11 performs two two-dimensional spectrum signals (frequency domain spectrum) transformed by the two-dimensional Fourier transform unit 2 for each combination of two teacher images in N teacher images. Signal) is multiplied by a complex conjugate to calculate a cross spectrum signal of two teacher images. The complex conjugate multiplication unit 11 constitutes a cross spectrum signal calculation unit.

The two-dimensional inverse Fourier transform unit 12 performs a two-dimensional inverse Fourier transform process on the cross spectrum signal calculated by the complex conjugate multiplication unit 11 for each combination of two teacher images, thereby converting the cross spectrum signal into a spatial domain. The process of converting to the signal is performed. The two-dimensional inverse Fourier transform unit 12 constitutes a spatial domain transform unit.
For each combination of two teacher images, the peak detection unit 13 detects the peak value of the signal in the spatial region converted by the two-dimensional inverse Fourier transform unit 12, and detects the peak position that is the position of the peak value. Implement the process.

The clustering unit 14 regards the peak value detected from the signal in the spatial region by the peak detection unit 13 as the fitness between two teacher images. If the fitness is higher than a preset threshold, the two teachers By repeating the process of classifying images into the same cluster, N teacher images are clustered.
Using the peak position detected by the peak detector 13, the parallel movement amount determiner 15 estimates the amount of misalignment between the teacher images classified into the same cluster by the clustering unit 14, and the clustering unit based on the amount of misalignment. 14 is performed to determine the parallel movement amount of the teacher images classified into the same cluster.
The teacher image translation unit 16 performs a process of translating the teacher image by the translation amount determined by the translation amount determination unit 15.

Next, the operation will be described.
When the teacher image data indicating N teacher images is input from the teacher image storage unit 1, the two-dimensional Fourier transform unit 2 performs the two-dimensional Fourier transform process on the N teacher image data to thereby obtain N teacher images. Each of the image data is converted into a spectrum signal in the frequency domain.
The two-dimensional Fourier transform unit 2 converts N pieces of teacher image data into a two-dimensional spectrum signal (frequency domain spectrum signal), raster-scans the two-dimensional spectrum signal, and vectorizes the two-dimensional spectrum signal. x _{1, x} 2, and outputs ..., and _{x N} to the teacher image registration unit 3.
In FIG. 2, teacher image data indicating four teacher images A, B, C, and D are input to the two-dimensional Fourier transform unit 2 and four vectorized two-dimensional spectrum signals x ₁ , x ₂ , x _3, x _N are output to the teacher image registration unit 3.

Upon receiving the vectorized two-dimensional spectrum signals x ₁ , x ₂ ,..., X _N from the two-dimensional Fourier transform unit 2, the teacher image alignment unit 3 receives the two-dimensional spectrum signals x ₁ , x ₂ , .., X _N is used to calculate the fitness between the N teacher images, and the N teacher images are classified into one or more clusters based on the fitness, and belong to the same cluster Perform alignment processing between teacher images.
In the example of FIG. 2, since the difference between the teacher images A, B, and D is small, they are classified into the same cluster. However, the teacher image C is different from the teacher images A, B, and D, so that another cluster is used. It is classified. For this reason, alignment processing is performed between the teacher images A, B, and D belonging to the same cluster. For the teacher image C, no other teacher images are classified in the same cluster, so that the alignment process is not performed.
Hereinafter, the teacher image alignment processing by the teacher image alignment unit 3 will be described in detail.

式（６）では、クロススペクトル信号ｚ_ｍ，ｎをｘ_ｍとｘ_ｎのエネルギーで正規化している。 When the complex conjugate multiplication unit 11 of the teacher image alignment unit 3 receives the two-dimensional spectrum signals x ₁ , x ₂ ,..., X _N from the two-dimensional Fourier transform unit 2, N teacher images are obtained. The cross spectrum signal of two teacher images is calculated by multiplying the complex conjugate of two two-dimensional spectrum signals (frequency domain spectrum signals) for each combination of two teacher images.
Equation (6) below shows a formula for calculating the cross spectrum signal z _{m, n} when the combination of two teacher images is the m-th teacher image and the n-th teacher image.

In Expression (6), the cross spectrum signal z _{m, n} is normalized by the energy of x _m and x _n .

When the complex conjugate multiplication unit 11 calculates the cross spectrum signal z _{m, n} for each combination of two teacher images, the two-dimensional inverse Fourier transform unit 12 calculates the cross spectrum signal z for each combination of two teacher images. _m, by carrying out a two-dimensional inverse Fourier transform process on _n, to convert the cross-spectrum signal z _{m, n} to the signal in the spatial domain.
It is known that the signal in the spatial domain corresponds to a cross-correlation function between the mth teacher image and the nth teacher image.

When the two-dimensional inverse Fourier transform unit 12 converts the cross spectrum signal z _{m, n} into a spatial domain signal for each combination of two teacher images, the peak detection unit 13 As the peak value of the signal in the spatial domain, the maximum value P _{m, n} (m = 1,..., N−1; n = m + 1,. ) Is detected.
The range of the maximum value P _{m, n} is 0 ≦ P _{m, n} ≦ 1, and can be used as an index representing the degree of matching between the m-th teacher image and the n-th teacher image.
When the peak detection unit 13 detects the maximum value P _{m, n} that is the fitness between the m-th teacher image and the n-th teacher image, the position of the cross-correlation function corresponding to the maximum value P _{m, n} (xy coordinates) ) Is detected.

When the peak detection unit 13 detects the maximum value P _{m, n} for each combination of two teacher images, the clustering unit 14 determines the maximum value P _{m, n} between the mth teacher image and the nth teacher image. The degree of conformity is regarded, and the degree of conformance _{Pm, n} is compared with a preset threshold value _Pth .
If the matching level P _{m, n} is greater than the threshold value P _{th , the} clustering unit 14 has a small difference between the m-th teacher image and the n-th teacher image. Classify to the same cluster.
The clustering unit 14 repeats the above process for all combinations of two teacher images, thereby clustering N teacher images.

As a result, the N teacher images are classified into several clusters. At this time, a teacher image that is not classified into the same cluster as other teacher images and is isolated may also be generated. In this case, a class identification filter is generated with one teacher image.
Hereinafter, Q-number of clusters generated by the clustering unit 14, to represent a set of numbers of teacher images included in the q-th cluster Omega _q.
In the example of FIG. 2, since the difference between the teacher images A, B, and D is small, they are classified into the same cluster. However, the teacher image C is different from the teacher images A, B, and D, so that another cluster is used. It is classified. As a result, the four teacher images A, B, C, and D are classified into two clusters.

The translation amount determination unit 15 uses the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} detected by the peak detection unit 13, and uses the peak position of the cross correlation function between the teacher images classified into the same cluster by the clustering unit 14. Estimate the amount of displacement.
For example, if the m-th teacher image and the n-th teacher image are classified into the same cluster, if the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} indicates the origin of the xy coordinates, It is determined that there is no displacement between the m-th teacher image and the n-th teacher image.
On the other hand, if the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} does not indicate the origin of the xy coordinates, there is a position shift between the m-th teacher image and the n-th teacher image, and the peak It is estimated that the x coordinate value of the position is the amount of positional deviation in the x direction, and the y coordinate value of the peak position is the amount of positional deviation in the y direction.

When the amount of positional deviation between the teacher images classified into the same cluster is estimated, the parallel movement amount determination unit 15 determines the amount of parallel movement of the teacher images classified into the same cluster from the amount of positional deviation.
For example, when three teacher images (teacher images A, B, and D) are classified into the same cluster, an arbitrary teacher image is selected from the teacher images A, B, and D as a reference image. If the reference image is the teacher image A, the translation amount of the teacher image A is 0, the translation amount of the teacher image B is the amount of misalignment between the teacher images A and B, and the translation of the teacher image D. The amount is determined like the amount of misalignment between the teacher image A and the teacher image D.

When the translation amount determination unit 15 determines the translation amount of the teacher images classified into the same cluster, the teacher image translation unit 16 converts the transformation matrix D _n (nεΩ) corresponding to the translation amount of the teacher image. _q ).
Note that the transformation matrix D _n (n∈Ω _q ) is a diagonal matrix, and the diagonal component corresponds to the positional deviation amount estimated by the positional deviation amount estimation unit 15 and is linear determined according to the Fourier transform shift rule. It is a complex signal having a correct phase component.

When the teacher image translation unit 16 generates a transformation matrix D _n (n∈Ω _q ) corresponding to the translation amount of the teacher image, the transformation is performed to align the positions of the teacher images classified into the same cluster. By multiplying the matrix D _n (nεΩ _q ) by the spectrum signal x _n (nεΩ _q ) of the teacher image, the teacher image is translated by the amount of translation.
The teacher image translation unit 16 outputs the two-dimensional spectrum signal Y of the teacher image after translation to the class identification filter synthesis unit 4.

In the first embodiment, the complex conjugate multiplication unit 11 of the teacher image registration unit 3 calculates cross spectrum signals of two teacher images for each combination of two teacher images in N teacher images. However, if an appropriate teacher image is selected in advance and only alignment of the teacher image is necessary, it is not essential to calculate cross-spectral signals for all two combinations. .
For example, it is possible to select one teacher image as a reference for the target position and calculate a cross spectrum between the teacher image and other N-1 teacher images.

ただし、

For each cluster classified by the teacher image registration unit 3, the class identification filter synthesis unit 4 substitutes the translated two-dimensional spectrum signal Y output from the teacher image registration unit 3 into the following equation (8). Then, a MINACE filter h that is a class identification filter is generated, and the class identification filter is stored in the class identification filter storage unit 5.
At this time, since the positions between the plurality of teacher images classified into the same cluster are accurately aligned, the problem that the shape of the class identification filter spreads as shown in FIG. 2 is avoided. The

However,

式（１０）において、Ｗは雑音の共分散行列であり、全ての教師画像の電力スペクトルの値が雑音レベルの定数倍（ｃ倍）を下回る場合、教師画像の電力スペクトルの代わりに、雑音レベルの定数倍（ｃ倍）の値で置き換える。
なお、式（１０）の最後のｃＷは、生成されるＭＩＮＡＣＥフィルタｈの雑音耐性を高めるために導入されている。 T is a diagonal matrix, and the diagonal component is given by the maximum value of the power spectrum of all the teacher images as shown in the following equation (10).

In Expression (10), W is a noise covariance matrix, and when the power spectrum values of all the teacher images are less than a constant multiple (c times) of the noise level, the noise level is used instead of the power spectrum of the teacher image. Replace with a value that is a constant multiple of (c times).
Note that the last cW in Expression (10) is introduced in order to increase the noise tolerance of the generated MINACE filter h.

Here, the MINACE filter h is given in the form of a “weighted average” of the vectorized two-dimensional spectrum signals D ₁ x ₁ , D ₂ x ₂ ,..., D _N x _N. It is defined in space.
Therefore, when the MINACE filter h is subjected to two-dimensional inverse Fourier transform, the shape of the target class identification filter in the image region is obtained as shown in the class identification filter of FIG.
Strictly speaking, since the MINACE filter h is given in the form of a vector, the two-dimensional spectrum signal x ₁ , It is necessary to perform a two-dimensional inverse Fourier transform after returning the MINACE filter h to a two-dimensional signal in the reverse manner of obtaining x ₂ ,..., x _N.

  In the example of FIG. 2, a class identification filter including three teacher images A, B, and D and a class identification filter including one teacher image C are generated, and two class identification filters are stored in the class identification filter. Stored in the unit 5.

  When the target type determination unit 6 inputs the observation image of the target of the class identification target, the target type determination unit 6 compares the observation image with one or more class identification filters stored in the class identification filter storage unit 5, A class identification filter having a matching degree equal to or higher than a preset threshold is searched, and the target type corresponding to the class identification filter is recognized as the target type of the class identification target.

  As is apparent from the above, according to the first embodiment, the degree of fitness between a plurality of teacher images is calculated using the frequency domain spectrum signal transformed by the two-dimensional Fourier transform unit 2, and the degree of fitness is calculated. Is provided with a teacher image alignment unit 3 that classifies a plurality of teacher images into one or more clusters and aligns the teacher images belonging to the same cluster. Since the class identification filter is generated by synthesizing the teacher images aligned by the teacher image alignment unit 3 for each cluster classified by the alignment unit 3, the amount of calculation is increased. Without any effect, it is possible to generate a class identification filter having a high target class identification performance.

That is, according to the first embodiment, before synthesizing a plurality of teacher images, the degree of matching between the teacher images is evaluated, and the teacher images having sufficiently high degrees of matching are classified into the same cluster. Therefore, it is possible to prevent the generation of a class identification filter having a low target identification performance by synthesizing an excessive number of teacher images.
In addition, before synthesizing the plurality of teacher images, the two-dimensional spectrum signal Y is synthesized after aligning the positions of the plurality of teacher images based on the positional deviation amount estimated from the peak position of the cross-correlation function between the teacher images. Thus, since the class identification filter is generated, the problem that the class identification filter spreads out can be avoided.
Furthermore, when calculating the cross-correlation function, the cross-spectrum calculation has been performed, and the alignment process between the teacher images is performed only by multiplying the phase function in the frequency domain. As compared with the alignment method described in No. 1, the calculation speed can be increased.

Embodiment 2. FIG.
FIG. 4 is a block diagram showing a target class identification device in which a class identification filter generation device according to Embodiment 2 of the present invention is mounted. In the figure, the same reference numerals as those in FIG. Omitted.
The teacher image alignment unit 20 is composed of, for example, a semiconductor integrated circuit on which a CPU is mounted, a one-chip microcomputer, or the like. Like the teacher image alignment unit 3 in FIG. Using the converted frequency domain spectral signal, the degree of matching between a plurality of teacher images is calculated, the plurality of teacher images are classified into one or more clusters based on the degree of matching, and belong to the same cluster The teacher image position shown in FIG. 1 is calculated in that the degree of matching between the plurality of teacher images is calculated in a state where the orientations of the targets existing in the plurality of teacher images are aligned. This is different from the mating unit 3. The teacher image alignment unit 20 constitutes a teacher image alignment unit.

In the example of FIG. 4, a teacher image storage unit 1, a two-dimensional Fourier transform unit 2, a teacher image alignment unit 20, a class identification filter synthesis unit 4, a class identification filter storage unit 5, and a target that are components of the target class identification device. Although each of the type determination units 6 is assumed to be configured with dedicated hardware, the target class identification device may be configured with a computer.
When the target class identification device is configured by a computer, the teacher image storage unit 1 and the class identification filter storage unit 5 are configured on the internal memory or the external memory of the computer, and the two-dimensional Fourier transform unit 2, the teacher image alignment The program describing the processing contents of the unit 20, the class identification filter synthesis unit 4 and the target type determination unit 6 is stored in the memory of the computer, and the CPU of the computer executes the program stored in the memory. do it.

FIG. 5 is a block diagram showing the teacher image alignment unit 20 of the class identification filter generation device according to Embodiment 2 of the present invention.
5, log-polar transformation unit 21 two-dimensional spectra output from the two-dimensional Fourier transform unit 2 signals x _1, x _2, · · ·, a x _N logarithmic polar coordinate transformation, log-polar transform is the result of the conversion Processing for outputting signals ζ ₁ , ζ ₂ ,..., Ζ _N is performed.
The two-dimensional inverse Fourier transform unit 22 log-polar transform signal zeta ₁ output from the logarithmic polar coordinate transformation unit _21, zeta _2, by implementing ..., the two-dimensional inverse Fourier transform processing on the zeta _N, the log-polar transform A process of converting the signals ζ ₁ , ζ ₂ ,..., Ζ _N into a spatial domain signal is performed.

The complex conjugate multiplier 23 performs a process of multiplying the complex conjugate of the two spatial domain signals transformed by the two-dimensional inverse Fourier transform unit 22 for each combination of two teacher images in the N teacher images. carry out.
The two-dimensional Fourier transform unit 24 performs a process of performing a two-dimensional Fourier transform on the multiplication result of the complex conjugate multiplication unit 23.
For each combination of two teacher images, the peak detection unit 25 detects the peak value of the two-dimensional Fourier transform result of the two-dimensional Fourier transform unit 24 and detects the peak position that is the position of the peak value. To do.
The rotation amount calculation unit 26 detects a positional deviation amount between the logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ ,..., Ζ _N output from the logarithmic polar coordinate conversion unit 21 from the peak position detected by the peak detection unit 25. Corresponding to the amount of deviation of the target direction existing in a plurality of teacher images and corresponding to the scaling and rotation amount between the plurality of teacher images), and logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ , ..., and it carries out a process of calculating respectively the amount of rotation of the plurality of teacher images from the amount of positional deviation between the zeta _N.
The two-dimensional inverse Fourier transform unit 22, the complex conjugate multiplication unit 23, the two-dimensional Fourier transform unit 24, the peak detection unit 25, and the rotation amount calculation unit 26 constitute a rotation amount calculation means.

The teacher image rotation unit 27 performs a process of rotating a plurality of teacher images by the rotation amount calculated by the rotation amount calculation unit 26.
The two-dimensional Fourier transform unit 28 performs a two-dimensional Fourier transform process on the teacher image data indicating the plurality of teacher images rotated by the teacher image rotation unit 27, thereby converting the plurality of teacher image data into spectrum signals in the frequency domain. Convert.
The teacher image rotation unit 27 and the two-dimensional Fourier transform unit 28 constitute a teacher image rotation unit.

  The complex conjugate multiplier 29 is a complex of two two-dimensional spectrum signals (frequency domain spectrum signals) transformed by the two-dimensional Fourier transform unit 28 for each combination of two teacher images of N teacher images. A process of calculating cross spectrum signals of two teacher images is performed by multiplying the conjugate. The complex conjugate multiplication unit 29 constitutes a cross spectrum signal calculation unit.

The two-dimensional inverse Fourier transform unit 30 performs a two-dimensional inverse Fourier transform process on the cross spectrum signal calculated by the complex conjugate multiplication unit 29 for each combination of two teacher images, thereby converting the cross spectrum signal into a spatial domain. The process of converting to the signal is performed. The two-dimensional inverse Fourier transform unit 30 constitutes a space domain transform unit.
For each combination of two teacher images, the peak detector 31 detects the peak value of the signal in the spatial region transformed by the two-dimensional inverse Fourier transform unit 30 and detects the peak position that is the position of the peak value. Implement the process.

The clustering unit 32 regards the peak value detected from the spatial domain signal by the peak detection unit 31 as the fitness between two teacher images, and if the fitness is higher than a preset threshold, the two teachers By repeating the process of classifying images into the same cluster, N teacher images are clustered.
Using the peak position detected by the peak detection unit 31, the parallel movement amount determination unit 33 estimates the amount of misalignment between the teacher images classified into the same cluster by the clustering unit 32, and the clustering unit based on the amount of misalignment. The process of determining the parallel movement amount of the teacher images classified into the same cluster by 32 is performed.
The teacher image translation unit 34 performs a process of translating the teacher image by the translation amount determined by the translation amount determination unit 33.

The two-dimensional inverse Fourier transform unit 35 performs two-dimensional inverse Fourier transform processing on the teacher image data indicating a plurality of teacher images translated by the teacher image translation unit 34, thereby converting the teacher image data into a spatial domain signal. Perform the process of converting to.
The teacher image reverse rotation unit 36 is a spatial region signal converted by the two-dimensional Fourier transform unit 35 by the rotation amount calculated by the rotation amount calculation unit 26 (a plurality of teacher images translated by the teacher image translation unit 34). The signal corresponding to 1) is rotated in the original direction.
The two-dimensional Fourier transform unit 37 performs a two-dimensional Fourier transform process on a plurality of spatial domain signals reversely rotated by the teacher image reverse rotation unit 36, thereby converting the spatial domain signals into frequency domain signals. To implement.
The two-dimensional inverse Fourier transform unit 35, the teacher image reverse rotation unit 36, and the two-dimensional Fourier transform unit 37 constitute teacher image reverse rotation means.

In the first embodiment, the teacher image alignment unit 3 uses the frequency domain spectrum signal transformed by the two-dimensional Fourier transform unit 2 to calculate the fitness between a plurality of teacher images. The degree of fitness between a plurality of teacher images is calculated without considering the direction of the target existing in the image.
For this reason, even if the orientations of the targets existing in the plurality of teacher images are different, the matching degree between the plurality of teacher images may be increased and classified into the same cluster. The target class identification performance by the class identification filter generated from the teacher image belonging to the cluster is degraded.
Therefore, in this second embodiment, when calculating the degree of fitness between a plurality of teacher images, if the orientations of the targets existing in the plurality of teacher images are different, the plurality of teacher images are rotated and a plurality of teacher images are rotated. The degree of matching between a plurality of teacher images is calculated in a state where the orientations of the targets existing in the teacher images are aligned.
However, in the case where the difference in target orientations existing in a plurality of teacher images is limited, the accuracy of calculation of the fitness level is not greatly affected. As shown in FIG. 1, it is more practical to calculate the degree of matching between a plurality of teacher images without considering the direction of the target existing in the plurality of teacher images.

Next, the operation will be described.
However, since the components other than the teacher image alignment unit 20 are the same as those in the first embodiment, only the processing content of the teacher image alignment unit 20 will be described here.
When the logarithmic polar coordinate conversion unit 21 of the teacher image alignment unit 20 receives the two-dimensional spectral signals x ₁ , x ₂ ,..., X _N from the two-dimensional Fourier transform unit 2, the two-dimensional spectral signal is received. x _{1, x} 2, · · ·, a _{x N} logarithmic polar coordinate transformation, log-polar transform signal zeta ₁ is a result of the conversion, zeta _2, · · ·, and outputs the zeta _N.
That is, the logarithmic polar coordinate converter 21, two-dimensional spectrum signal x ₁ is an amplitude _spectrum, x _2, · · ·, a x _N logarithmic transformation to convert the log-polar in emphasized form the high frequency region.

When the two-dimensional inverse Fourier transform unit 22 receives the logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ ,..., Ζ _N from the logarithmic polar coordinate conversion unit 21, the logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ ,. By performing a two-dimensional inverse Fourier transform process on ζ _N , the logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ ,..., ζ _N are converted into spatial domain signals.
The complex conjugate multiplier 23 multiplies the complex conjugate of the two spatial domain signals transformed by the two-dimensional inverse Fourier transform unit 22 for each combination of two teacher images in the N teacher images.
The two-dimensional Fourier transform unit 24 performs a two-dimensional Fourier transform on the multiplication result of the complex conjugate multiplication unit 23.

The peak detection unit 25 detects the peak value of the two-dimensional Fourier transform result of the two-dimensional Fourier transform unit 24 for each combination of two teacher images and also detects the peak position that is the position of the peak value.
That is, for each combination of two teacher images, the peak detection unit 25 uses the maximum absolute value of the cross-correlation function of the two teacher images as the peak value of the two-dimensional Fourier transform result of the two-dimensional Fourier transform unit 24. P _{m, n} (m = 1,..., N−1; n = m + 1,..., N) is detected.
When the peak detection unit 25 detects the maximum value P _{m, n} that is the fitness between the m-th teacher image and the n-th teacher image, the position of the cross-correlation function corresponding to the maximum value P _{m, n} (xy coordinates) ) Is detected.

When the peak detection unit 25 detects the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} for each combination of two teacher images, the rotation amount calculation unit 26 performs logarithmic polar coordinate conversion using the peak position. The amount of positional deviation between the logarithmic polar coordinate conversion signal ζ _m output from the unit 21 and the logarithmic polar coordinate conversion signal ζ _n is estimated. The positional deviation amount between the logarithmic polar coordinate conversion signal ζ _m and the logarithmic polar coordinate transformation signal ζ _n corresponds to the deviation amount of the target direction existing in the mth teacher image and the nth teacher image.
For example, if the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} does not indicate the origin of the xy coordinates, the amount of deviation in the target direction existing in the m-th teacher image and the n-th teacher image It is estimated that the x coordinate value of the peak position is the deviation amount of the target direction.

When calculating the amount of positional deviation between the logarithmic polar coordinate conversion signals ζ ₁ , ζ ₂ ,..., Ζ _N , for example, the rotation amount calculating unit 26 multiplies the positional deviation amount by a preset coefficient. Thus, the rotation amount (rotation amount = position displacement amount × coefficient) of the N teacher images is calculated.
For example, when there are three teacher images (teacher images A, B, and D), an arbitrary teacher image is selected from the teacher images A, B, and D as a reference image. If the reference image is the teacher image A, the rotation amount of the teacher image A is 0, the rotation amount of the teacher image B is the displacement amount of the teacher image A and the teacher image B × coefficient, and the rotation amount of the teacher image D Is determined as follows: Misalignment between teacher image A and teacher image D × coefficient.

When the rotation amount calculation unit 26 calculates the rotation amount of the N teacher images, the teacher image rotation unit 27 rotates the N teacher images by the rotation amount.
The two-dimensional Fourier transform unit 28 performs a two-dimensional Fourier transform process on the teacher image data indicating the N teacher images rotated by the teacher image rotation unit 27, thereby converting the N teacher image data into a frequency domain spectrum. Convert each into a signal.
When the two-dimensional Fourier transform unit 28 converts N pieces of teacher image data into a two-dimensional spectrum signal (frequency domain spectrum signal), the two-dimensional spectrum signal is raster-scanned and vectorized, and the vectorized two-dimensional spectrum signal is obtained. x _{1, x} 2, and outputs ..., and _{x N} complex conjugate multiplying unit 29.
Here, the two-dimensional spectrum signal output from the two-dimensional Fourier transform unit 28 to the complex conjugate multiplier 29 is used as the two-dimensional spectrum output from the two-dimensional Fourier transform unit 2 to the complex conjugate multiplier 11 in the first embodiment. The two-dimensional spectrum signal output from the two-dimensional Fourier transform unit 28 to the complex conjugate multiplication unit 29 is represented by the same symbol (x ₁ , x ₂ ,..., X _N ) as the signal. The difference is that the signal is obtained from a teacher image in which the target orientation is aligned.

When the complex conjugate multiplier 29 receives the vectorized two-dimensional spectrum signals x ₁ , x ₂ ,..., X _N from the two-dimensional Fourier transform unit 28, as in the complex conjugate multiplier 11 of FIG. Cross spectrum signals of two teacher images by multiplying a complex conjugate of two two-dimensional spectrum signals (frequency domain spectrum signals) for each combination of two teacher images in N teacher images. Is calculated.
The cross spectrum signal z _{m, n} when the combination of the two teacher images is the m-th teacher image and the n-th teacher image is expressed by the above equation (6).

When the complex conjugate multiplication unit 29 calculates the cross spectrum signal z _{m, n} for each combination of two teacher images, the two-dimensional inverse Fourier transform unit 30 is similar to the two-dimensional inverse Fourier transform unit 12 in FIG. By performing a two-dimensional inverse Fourier transform process on the cross spectrum signal z _{m, n} for each combination of teacher images, the cross spectrum signal z _{m, n} is converted into a spatial domain signal.
It is known that the signal in the spatial domain corresponds to a cross-correlation function between the mth teacher image and the nth teacher image.

When the two-dimensional inverse Fourier transform unit 30 converts the cross spectrum signal z _{m, n} into a spatial domain signal for each combination of two teacher images, the peak detection unit 31 is similar to the peak detection unit 13 of FIG. For each combination of two teacher images, the maximum value P _{m, n} (m = 1,..., N− 1; n = m + 1,..., N) are detected.
The range of the maximum value P _{m, n} is 0 ≦ P _{m, n} ≦ 1, and can be used as an index representing the degree of matching between the m-th teacher image and the n-th teacher image.
When the peak detection unit 31 detects the maximum value P _{m, n} that is the fitness between the m-th teacher image and the n-th teacher image, the position of the cross-correlation function corresponding to the maximum value P _{m, n} (xy coordinates) ) Is detected.

The clustering unit 32, the maximum value for each combination of the peak detector 31 are two teacher image P _m, detects a _n, similarly to the clustering unit 14 of FIG. 3, the maximum value P _m, the _n of the m-th The degree of matching between the teacher image and the nth teacher image is considered, and the degree of matching P _{m, n} is compared with a preset threshold value P _th .
Since the difference between the m-th teacher image and the n-th teacher image is small if the matching degree P _{m, n} is larger than the threshold value P _{th , the} clustering unit 32 determines the m-th teacher image and the n-th teacher image. Classify to the same cluster.
The clustering unit 32 clusters the N teacher images by repeating the above processing for all combinations of the two teacher images.

Similar to the parallel movement amount determination unit 15 in FIG. 3, the parallel movement amount determination unit 33 uses the peak position of the cross-correlation function corresponding to the maximum value P _{m, n} detected by the peak detection unit 31 to use the clustering unit. 32, the amount of misalignment between the teacher images classified into the same cluster is estimated.
When the amount of positional deviation between the teacher images classified into the same cluster is estimated, the parallel movement amount determination unit 33 determines the amount of parallel movement of the teacher images classified into the same cluster from the amount of positional deviation.

When the translation amount determination unit 33 determines the translation amount of the teacher image classified into the same cluster, the teacher image translation unit 34 determines that the teacher image parallel unit 34 is parallel to the teacher image. A transformation matrix D _n (nεΩ _q ) corresponding to the movement amount is generated.
When the teacher image translation unit 34 generates the transformation matrix D _n (n∈Ω _q ) corresponding to the translation amount of the teacher image, the transformation is performed in order to align the positions of the teacher images classified into the same cluster. By multiplying the matrix D _n (nεΩ _q ) by the spectrum signal x _n (nεΩ _q ) of the teacher image, the teacher image is translated by the amount of translation.
The teacher image translation unit 34 outputs the two-dimensional spectrum signal Y of the teacher image after translation to the class identification filter synthesis unit 4.

When the two-dimensional inverse Fourier transform unit 35 receives the two-dimensional spectrum signal Y of the teacher image after translation from the teacher image translation unit 34, the two-dimensional inverse Fourier transform unit 35 performs a two-dimensional inverse Fourier transform process on the two-dimensional spectrum signal Y. Convert to spatial domain signal.
The teacher image reverse rotation unit 36 is a signal of the spatial region converted by the two-dimensional Fourier transform unit 35 by a rotation amount calculated by the rotation amount calculation unit 26 (a plurality of teachers translated by the teacher image translation unit 34). The signal corresponding to the image) is rotated in the original direction, thereby obtaining an image obtained by translating the original teacher image by the estimated positional deviation.
The two-dimensional Fourier transform unit 37 performs a two-dimensional Fourier transform process on a plurality of spatial domain signals reversely rotated by the teacher image reverse rotation unit 36, thereby converting the spatial domain signals into frequency domain signals. .

As apparent from the above, according to the second embodiment, the degree of fitness between the plurality of teacher images in a state where the teacher image alignment unit 20 matches the orientations of the targets existing in the plurality of teacher images. Therefore, even when the orientations of targets existing in a plurality of teacher images are different, it is possible to generate a class identification filter with high target class identification performance.
In addition, according to the second embodiment, logarithmic polar coordinate conversion is used, and the rotation amount between different teacher images is estimated by estimating the shift amount of the signal after logarithmic polar coordinate conversion. Can be kept low. As a result, it is possible to cluster teacher images having a high degree of fitness without excess or deficiency, and thus it is possible to prevent a reduction in target identification performance.

Embodiment 3 FIG.
FIG. 6 is a block diagram showing the teacher image alignment unit 20 of the class identification filter generating device according to Embodiment 3 of the present invention. In the figure, the same reference numerals as those in FIG. .
The image position / rotation fine adjustment unit 38 finely adjusts the position and orientation with the fine adjustment amount while changing the fine adjustment amount of the position and orientation of the teacher image belonging to the same cluster translated by the teacher image translation unit 34. In this case, the entropy is calculated, the fine adjustment amount that minimizes the entropy is specified, and the position and orientation of the teacher images belonging to the same cluster are adjusted by the fine adjustment amount. The image position / rotation fine adjustment unit 38 constitutes an adjustment unit.

In the second embodiment, the teacher image alignment unit 20 calculates the fitness between the plurality of teacher images in the state where the orientations of the targets existing in the plurality of teacher images are aligned. In the radar observation image, when the target rotates, the influence of the layover and the phase center of reflection change, so that the target image may not match even if the teacher image is simply rotated. If the target image matches when the teacher image is rotated, it is possible to accurately estimate the rotation amount of the teacher image, but the target image does not completely match even if the teacher image is rotated In some cases, the estimation accuracy of the rotation amount may deteriorate.
Therefore, in the third embodiment, an image position / rotation fine adjustment unit 38 that finely adjusts the position and orientation of the teacher image is provided after the teacher image is rotated and translated.

式（１２）では、ｑ番目のクラスタに分類されている教師画像について示しており、ｎ∈Ω_ｑ番目の教師画像を微小回転量Δθだけ回転させて、微小平行移動量（Δｘ，Δｙ）平行移動させていることを表している。
ただし、式（１２）の処理においては、ｑ番目のクラスタに分類されている教師画像のいずれかを基準に設定しており、基準に設定した教師画像については回転も移動もさせていない。即ち、基準画像については、常に（Δθ，Δｘ，Δｙ）＝（０，０，０）である。 The processing contents of the image position / rotation fine adjustment unit 38 will be specifically described below.
When the image position / rotation fine adjustment unit 38 receives the teacher image data ζ _n indicating the teacher image rotated and translated from the teacher image from the two-dimensional inverse Fourier transform unit 35, the image position / rotation fine adjustment unit 38 performs various processing on the teacher image data ζ _n . A small rotation amount Δθ and a small parallel movement amount (Δx, Δy) are assumed.
Assuming various minute rotation amounts Δθ and minute parallel movement amounts (Δx, Δy), the image position / rotation fine adjustment unit 38 assumes the assumed minute rotation amount Δθ and minute amount as shown in the following equation (12). Alignment of each teacher image after rotation and translation is performed with the translation amount (Δx, Δy).

Equation (12) shows the teacher image classified into the q-th cluster, and the nεΩ _q- th teacher image is rotated by the minute rotation amount Δθ to be parallel to the minute parallel movement amount (Δx, Δy). Indicates that it is moved.
However, in the processing of Expression (12), one of the teacher images classified into the qth cluster is set as a reference, and the teacher image set as the reference is neither rotated nor moved. That is, for the reference image, (Δθ, Δx, Δy) = (0, 0, 0) is always established.

式（１３）で与えられる共分散行列Ｒ_Ωｑの固有値をλ_ｉとすると、エントロピーＥは、下記の式（１４）で定義される。

The image position / rotation fine adjustment unit 38 performs alignment using not the cross-correlation function but the entropy E defined below as an index.
That is, the image position / rotation fine adjustment unit 38 assumes various micro rotation amounts Δθ and micro translation amounts (Δx, Δy), and adds χ _n , (n∈Ω _q ) To define the following covariance matrix: χ _n is a vector obtained by raster scanning.

When the eigenvalue of the covariance matrix R _Ωq given by the equation (13) is λ _i , the entropy E is defined by the following equation (14).

The image position / rotation fine adjustment unit 38 calculates entropy E for all combinations of χ _n , (n∈Ω _q ), and the minute rotation amount and minute translation amount (Δθ, Δx, to explore the combination of Δy) n∈Ω _q.
When the image position / rotation fine adjustment unit 38 searches for a combination of a minute rotation amount and a minute translation amount (Δθ, Δx, Δy) nεΩ _q that minimizes entropy E, the minute rotation amount and minute translation amount. (Δθ, Δx, Δy) nεΩ _q is used to finely adjust the orientation and position of the teacher image classified in the cluster, and the teacher image data indicating the teacher image after the fine adjustment is supplied to the teacher image reverse rotation unit 36. Output.

  As is apparent from the above, according to the third embodiment, the image position / rotation fine adjustment unit 38 is used to finely adjust the position and orientation of the teacher images belonging to the same cluster translated by the teacher image translation unit 34. While changing the adjustment amount, calculate the entropy when the position and orientation are fine-adjusted with the fine adjustment amount, identify the fine adjustment amount that minimizes the entropy, and use the fine adjustment amount to identify teachers belonging to the same cluster Since it is configured to adjust the position and orientation of the image, even if image distortion occurs due to the influence of layover or the phase center of reflection, the estimation error of the rotation amount due to the decrease in cross-correlation It is possible to reduce the influence and prevent the target identification performance from being lowered.

  In the third embodiment, the teacher image registration unit 20 calculates entropy E when the position and orientation are finely adjusted with various assumed fine adjustment amounts, and the fine adjustment amount that minimizes the entropy E is calculated. It was specified that the position and orientation of the teacher images belonging to the same cluster are adjusted with the fine adjustment amount, but the mutual information amount (optical) when the position and orientation are finely adjusted with various assumed fine adjustment amounts The amount of information that is generally used for image alignment, which can be an amount that represents a measure of the interdependence of two random variables), and the amount of fine adjustment that maximizes the amount of mutual information The position and orientation of the teacher images belonging to the same cluster may be adjusted with the fine adjustment amount, and the same effect can be obtained.

  In the present invention, within the scope of the invention, any combination of the embodiments, or any modification of any component in each embodiment, or omission of any component in each embodiment is possible. .

  DESCRIPTION OF SYMBOLS 1 Teacher image storage part (teacher image storage means), 2 Two-dimensional Fourier-transform part (frequency domain conversion means), 3 Teacher image position alignment part (teacher image position alignment means), 4 Class identification filter composition part (Class identification filter composition) Means) 5 type identification filter storage unit, 6 target type determination unit (target type determination unit), 11 complex conjugate multiplication unit (cross spectrum signal calculation unit), 12 two-dimensional inverse Fourier transform unit (spatial domain conversion unit), 13 Peak detection unit, 14 clustering unit, 15 translation amount determination unit, 16 teacher image translation unit, 20 teacher image alignment unit (teacher image alignment unit), 21 log polar coordinate conversion unit, 22 two-dimensional inverse Fourier transform unit ( Rotation amount calculation means), 23 complex conjugate multiplication section (rotation amount calculation means), 24 two-dimensional Fourier transform section (rotation amount calculation means), 25 peak detection Unit (rotation amount calculation unit), 26 rotation amount calculation unit (rotation amount calculation unit), 27 teacher image rotation unit (teacher image rotation unit), 28 two-dimensional Fourier transform unit (teacher image rotation unit), 29 complex conjugate multiplication unit (Cross spectrum signal calculation unit), 30 two-dimensional inverse Fourier transform unit (spatial domain transform unit), 31 peak detection unit, 32 clustering unit, 33 translation amount determination unit, 34 teacher image translation unit, 35 two-dimensional inverse Fourier Conversion unit (teacher image reverse rotation unit), 36 Teacher image reverse rotation unit (teacher image reverse rotation unit), 37 Two-dimensional Fourier transform unit (teacher image reverse rotation unit), 38 Image position / rotation fine adjustment unit (adjustment unit) , 101 teacher image data file, 102 two-dimensional Fourier transform means, 103 type identification filter synthesis means, 104 type identification filter.

Claims (7)

Teacher image storage means for storing a target teacher image cut out from a plurality of observation images of the target observed under different observation conditions;
Frequency domain conversion means for converting teacher image data indicating a plurality of teacher images stored by the teacher image storage means, respectively, into spectrum signals in the frequency domain;
Using the frequency domain spectrum signal transformed by the frequency domain transforming means, the fitness between the plurality of teacher images is calculated, and the teacher images are classified into one or more clusters based on the fitness. And teacher image alignment means for aligning teacher images belonging to the same cluster,
A class identification filter synthesizing unit that generates a class identification filter by synthesizing the teacher images aligned by the teacher image alignment unit for each cluster classified by the teacher image alignment unit; Identification filter generation device. The teacher image alignment means includes:
A cross spectrum for calculating a cross spectrum signal of two teacher images using a frequency domain spectrum signal converted by the frequency domain conversion means for each combination of two teacher images of the plurality of teacher images. A signal calculator;
A spatial domain conversion unit that converts the cross spectrum signal calculated by the cross spectrum signal calculation unit into a spatial domain signal for each combination of two teacher images;
For each combination of two teacher images, a peak detection unit that detects a peak value of a signal of the spatial region converted by the spatial region conversion unit and detects a peak position that is a position of the peak value;
The peak value detected from the signal in the spatial region by the peak detection unit is regarded as the fitness between two teacher images. If the fitness is higher than a preset threshold, the two teacher images are identical. A clustering unit for clustering the plurality of teacher images by classifying them into clusters,
Using the peak positions detected by the peak detection unit, the amount of misalignment between teacher images classified into the same cluster by the clustering unit is estimated, and the clustering unit classifies the misalignment amount from the position deviation amount into the same cluster. A translation amount determining unit for determining a translation amount of the teacher image that has been
The class identification filter generation device according to claim 1, further comprising: a teacher image translation unit that translates the teacher image by a translation amount determined by the translation amount determination unit.   2. The class according to claim 1, wherein the teacher image alignment unit calculates a degree of matching between the plurality of teacher images in a state where the orientations of targets existing in the plurality of teacher images are aligned. Identification filter generation device. The teacher image alignment means includes:
A logarithmic polar coordinate conversion unit for logarithmic polar coordinate conversion of each spectrum signal converted by the frequency domain conversion means, and outputting a logarithmic polar coordinate conversion signal as a result of the conversion;
From each logarithmic polar coordinate conversion signal output from the logarithmic polar coordinate conversion unit, a deviation amount of the target direction existing in the plurality of teacher images is estimated, respectively, and the rotation amount of the plurality of teacher images is calculated from the deviation amount. Rotation amount calculating means for calculating
Teacher image rotation means for rotating each of the plurality of teacher images by the rotation amount calculated by the rotation amount calculation means;
For each combination of two teacher images of the plurality of teacher images rotated by the teacher image rotating means, the frequency domain spectrum signal converted by the frequency domain converting means is used to generate two teacher images. A cross spectrum signal calculation unit for calculating a cross spectrum signal;
A spatial domain conversion unit that converts the cross spectrum signal calculated by the cross spectrum signal calculation unit into a spatial domain signal for each combination of two teacher images;
For each combination of two teacher images, a peak detection unit that detects a peak value of a signal of the spatial region converted by the spatial region conversion unit and detects a peak position that is a position of the peak value;
The peak value detected from the signal in the spatial region by the peak detection unit is regarded as the fitness between two teacher images. If the fitness is higher than a preset threshold, the two teacher images are identical. A clustering unit for clustering the plurality of teacher images by classifying them into clusters,
Using the peak positions detected by the peak detection unit, the amount of misalignment between teacher images classified into the same cluster by the clustering unit is estimated, and the clustering unit classifies the misalignment amount from the position deviation amount into the same cluster. A translation amount determining unit for determining a translation amount of the teacher image that has been
A teacher image translation unit that translates the teacher image by the translation amount determined by the translation amount determination unit;
A plurality of teacher images translated by the teacher image translation unit by a rotation amount calculated by the rotation amount calculation unit, respectively, and a teacher image reverse rotation unit that reversely rotates the teacher images in their original directions. The class identification filter generation device according to claim 3, wherein:   The teacher image alignment means is an entropy when the position and orientation are adjusted by the adjustment amount while changing the adjustment amount of the position and orientation of the teacher image belonging to the same cluster translated by the teacher image translation unit. 5. The class identification according to claim 4, further comprising: an adjustment unit that calculates the position of the teacher image by adjusting the position and orientation of the teacher image by specifying the adjustment amount that minimizes the entropy. Filter generation device.   The teacher image alignment unit is configured to perform mutual adjustment when the position and orientation are adjusted by the adjustment amount while changing the adjustment amount of the position and orientation of the teacher image belonging to the same cluster translated by the teacher image translation unit. 5. The apparatus according to claim 4, further comprising: an adjustment unit that calculates an information amount, specifies the adjustment amount that maximizes the mutual information amount, and adjusts the position and orientation of the teacher image by the adjustment amount. The kind identification filter production | generation apparatus of description. Teacher image storage means for storing a target teacher image cut out from a plurality of observation images of the target observed under different observation conditions;
Frequency domain conversion means for converting teacher image data indicating a plurality of teacher images stored by the teacher image storage means, respectively, into spectrum signals in the frequency domain;
Using the frequency domain spectrum signal transformed by the frequency domain transforming means, the fitness between the plurality of teacher images is calculated, and the teacher images are classified into one or more clusters based on the fitness. And teacher image alignment means for aligning teacher images belonging to the same cluster,
A class identification filter synthesizing unit that generates a class identification filter by synthesizing the teacher images registered by the teacher image alignment unit for each cluster classified by the teacher image alignment unit;
A target class identification apparatus comprising: a target type determination unit that compares an observed image of a target to be identified with a class identification filter generated by the class identification filter synthesis unit to determine the type of the target.

Images