JP2012127920A - Target type identification apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a target type identification apparatus which is capable of dealing with a discrete distribution situation of observation element parameters in type identification filter generation merging a plurality of teacher images of different observation elements or the like.SOLUTION: A target type identification apparatus includes an image preprocessing section 3 which performs preprocessing prior to filter generation such as alignment on teacher image data of an observation target, which are obtained under a plurality of observation conditions, stored in a teacher image data file 1, a teacher image expansion section 4 which expands a high luminance region, for compensating for a discrete distribution of observation elements, for preprocessed teacher image data, a kind determination filter generation section 5 which generates a kind determination filter from two or more expanded teacher image data items, and an image collation section 6 which collates observation image data of a kind determination target stored in an observation image data file 2 with the kind determination filter generated by the kind determination filter generation section 5, makes a kind determination of the observation target, and stores a determination result in a kind determination result file 7.

Description

  The present invention relates to a target class identification device, and more particularly to a target class identification device for determining a target type for radar observation image data of an unknown target.

  Conventional methods for identifying target types include pattern matching with observation images of static targets obtained from radar observations with teacher images of each model obtained in advance to determine the target type. The method is known. The teacher image is generally either a simulation image (specifications are known) or an actual image (some specifications are unknown).

  In the method described in Non-Patent Document 1, a plurality of images with different observation specifications (target main axis direction (direction)) for the same target are integrated to generate a class identification filter.

  As preprocessing for filter integration, processing such as brightness adjustment, size normalization, and center alignment is performed.

  The method described in Non-Patent Document 1 can deal with images of different orientations by integrating images of various orientations without estimating parameters such as the target spindle direction, and is not matching with teacher images. By performing matching with a filter, there is a feature that performance does not depend on selection of a teacher image.

Rajesh K. Shenoy, “Object detection and classification in SAR images using MINACE correlation filters”, A project report for the degree of Master of Science, Carnegie Mellon University, Department of Electrical and Computer Engineering, April 10,1995

  In the conventional target class identification device, when a filter is generated between images having different target principal axis directions, in principle, a filter is generated by combining images obtained by discretely rotating the target.

  FIG. 2 shows an example of filter generation using two teacher images with different target spindle directions in the prior art. A blackish circle and a whitish circle filled with dots represent a high luminance point (hereinafter, a blackish circle is referred to as a black circle and a whitish circle is referred to as a white circle. The black circle is assumed to have a higher luminance value than the white circle). A type determination filter (FIG. 2 (c)) for the target A is generated from the teacher image 1 (FIG. 2 (a)) and the teacher image 2 (FIG. 2 (b)) having different target spindle directions with respect to the target A. In other words, the target A type determination filter shown in FIG. 2C is a combination of the teacher images 1 and 2 shown in FIGS.

  In the conventional method, when images of various target main axis directions (directions) are not obtained at the time of filter generation, there is a large difference in target main axis direction (direction) between the teacher image used for filter generation and the observed image that is the type determination target. In this case, the filter and the distribution status are greatly different and cannot be handled. Even if the target main axis direction difference is small, the filter generated based on the discrete distribution points tends to have a lower degree of conformity to the filter as the target main axis direction difference between the teacher image and the observation image increases. There was a problem of being.

  In addition to the target main axis direction, observation date and time, incident angle of radio waves transmitted from a radar mounted on an aircraft / satellite, polarization, resolution, etc. can be considered as observation parameters. Among these images with different observation specifications, especially images with a large difference in incident angle, the luminance distribution (luminance value, distribution position) and the like of the radar observation image differ due to a phenomenon called collapse, even in the same model.

  FIGS. 3 to 4 show examples of filter generation using two teacher images with different incident angles in the prior art. FIG. 3 is an example of filter generation using two teacher images with a relatively small incident angle difference, and FIG. 4 is an example of filter generation using two teacher images with a relatively large incident angle difference. 3 and 4, black circles and white circles represent high luminance points (black circles have higher luminance values than white circles).

  In FIG. 3, a type determination filter (FIG. 3C) of the target B is generated from the teacher image 1 (FIG. 3A) and the teacher image 2 (FIG. 3B) having different incident angles with respect to the target B. . Similarly, in FIG. 4, a target B type determination filter (FIG. 4C) is obtained from the teacher image 3 (FIG. 4A) and the teacher image 4 (FIG. 4B) having different incident angles with respect to the target B. Generated.

  When filter generation is performed between images with different incident angles, the luminance distribution diverges, particularly when the incident angle difference is relatively large as shown in FIG. False detection is likely to occur. In addition, as shown in FIG. 3, even in a filter generated between images having a relatively small incident angle difference, the incident angles are distributed discretely and the luminance distribution fluctuates accordingly. There was a problem that the degree of conformity to the filter tends to decrease as the angular difference increases.

  The present invention has been made to solve such a problem, and compensates for the influence of the discrete distribution of observation parameters and generates a filter with a high degree of fitness to obtain a highly accurate target class identification device. With the goal.

  The present invention is a target class identification device for determining the type of an observation target for observation image data of an observation target, and comprises teacher image data of the observation target obtained by observation under a plurality of observation conditions. A teacher image data file for storing a teacher image data group and its observation specification data, an observation image data to be subjected to type determination and an observation image data file for storing the observation specification data, and the teacher image data, An image pre-processing unit that performs pre-processing before filter generation, such as alignment, and generates pre-processed teacher image data after pre-processing, and the pre-processed teacher image data for the pre-processed teacher image data A teacher image extension unit that generates an extended teacher image data by extending a high-intensity region to compensate for the discrete distribution of observation parameters, and a type determination file based on the extended teacher image data. A type determination filter generation unit that generates a data, an image verification unit that performs a comparison between the observation image data and the type determination filter, and performs a type determination of the observation target, and a type that stores the type determination result of the observation target A target class identification device comprising a determination result file.

  The present invention is a target class identification device for determining the type of an observation target for observation image data of an observation target, and comprises teacher image data of the observation target obtained by observation under a plurality of observation conditions. A teacher image data file for storing a teacher image data group and its observation specification data, an observation image data to be subjected to type determination and an observation image data file for storing the observation specification data, and the teacher image data, An image pre-processing unit that performs pre-processing before filter generation, such as alignment, and generates pre-processed teacher image data after pre-processing, and the pre-processed teacher image data for the pre-processed teacher image data A teacher image extension unit that generates an extended teacher image data by extending a high-intensity region to compensate for the discrete distribution of observation parameters, and a type determination file based on the extended teacher image data. A type determination filter generation unit that generates a data, an image verification unit that performs a comparison between the observation image data and the type determination filter, and performs a type determination of the observation target, and a type that stores the type determination result of the observation target Since it is a target class identification device characterized by having a judgment result file, it compensates for the influence of the discrete distribution of observation parameters, generates a filter with high fitness, and performs high-precision type judgment Make it possible.

It is a block diagram which shows the structure of the target class identification device which concerns on Embodiment 1 of this invention. It is explanatory drawing which shows the type determination filter production | generation example by a prior art. It is explanatory drawing which shows the type determination filter production | generation example by a prior art. It is explanatory drawing which shows the type determination filter production | generation example by a prior art. It is explanatory drawing which shows the type determination filter production | generation example by Embodiment 1 of this invention. It is explanatory drawing which shows the type determination filter production | generation example by Embodiment 2 of this invention. It is explanatory drawing which shows the type determination filter production | generation example by Embodiment 3 of this invention. It is explanatory drawing which shows the type determination filter production | generation example by Embodiment 3 of this invention.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing the configuration of the target class identification device according to Embodiment 1 of the present invention. As shown in FIG. 1, the target class identification device according to the first embodiment stores a teacher image data group, which is a plurality of target teacher image data obtained under various observation conditions, and its observation data. Perform pre-processing before filter generation such as alignment for the teacher image data file 1, the observation image data file 2 that stores the observation image data to be subjected to type determination and the observation specification data, and the teacher image data group , An image pre-processing unit 3 that generates a pre-processed pre-processed teacher image data group, and an extension to supplement the discrete distribution of observation data of the teacher image data group with respect to the pre-processed teacher image data group And a teacher image extension unit 4 that generates a post-expansion teacher image data group, a type determination filter generation unit 5 that generates a type determination filter from the post-extension teacher image data group, and observation image data that is a type determination target Matches with a kind determining filter includes an image matching unit 6 that performs type determination observation target, and the type determination result file 7 for storing the type determination result of the observation target.

  Although not shown in FIG. 1, an imaging device and a signal processing device are connected to the input side of the observation image data file 2. The signal processing device calculates observation image data T based on the observation image captured by the imaging device and stores it in the observation image data file 2. In the observation image data file 2, the observation specification data of the observation image data T is also stored. The observation data includes, for example, the incident angle of the observation radar's transmitted radio wave at the time of observation, the direction of the observation radar's line of sight, the direction of the target main axis, the observation date and time, the observation time, the observation area, the resolution of the observation radar (Multi-polarization observation, etc.).

  The teacher image data file 1 is obtained by observing a target under various observation conditions in advance, and is stored in a computer or the like, known type of target image data, or each target generated by computer simulation. The teacher image data L consisting of the image data is stored. The teacher image data file 1 also stores observation data of the teacher image data L at the same time.

  The type determination result file 7 stores type determination result data R obtained by the image collation unit 6 described later.

  The image preprocessing unit 3 uses the teacher image data L from the teacher image data file 1 as input information, performs preprocessing such as brightness adjustment, size normalization, center alignment, etc. on the teacher image data L, The preprocessed image is output as preprocessed image data P.

  The teacher image extension unit 4 extends the high luminance point region with respect to the preprocessed image data P output from the image preprocessing unit 3, and generates post-extension teacher image data K.

  The type determination filter generation unit 5 generates a type determination filter F from the expanded teacher image data K output from the teacher image extension unit 4.

  The image collation unit 6 performs type determination of the observation image data T stored in the observation image data file 2 using the type determination filter F output from the type determination filter generation unit 5, and performs a type determination result R (total Output information) is generated and stored in the type determination result file 7.

  Next, the operations of the functional blocks 3 to 6 shown in FIG. 1 will be described more specifically with reference to the explanatory diagram of FIG.

  First, the image preprocessing unit 3 reads the teacher image data L from the teacher image data file 1 and performs luminance adjustment by, for example, the LOG method on the teacher image data L as a target. The LOG method is a method of reducing a luminance difference (shading difference, amplitude difference) by taking a logarithm of the luminance of an image, and improves the visibility of the image. Next, a scale adjustment process is performed based on the observation parameters of each image obtained in advance so that the target size per pixel does not depend on the observation parameters. That is, size normalization is performed. Furthermore, as the center alignment, a reference image, which is a reference image, is arbitrarily selected from the teacher images, and the reference image and each teacher image are overlaid while being shifted one pixel at a time. And the point where the correlation coefficient between the teacher image and the teacher image is maximized. These processes are collectively referred to as preprocessing. The image preprocessing unit 3 outputs the image data obtained by the preprocessing for the teacher image data L as preprocessed image data P.

  Subsequently, in the teacher image expansion unit 4, each high-luminance point is set for the preprocessed image data P output from the image preprocessing unit 3 in order to supplement the discrete distribution of the observation data of the teacher image data. The high brightness area is expanded at the center. The expansion of the high luminance area is performed by changing the luminance values of N neighboring pixels in the vicinity of the central high luminance point C to the same luminance value as that of the high luminance point C, or maximizing the luminance value of the high luminance point C. As a value, there is a method of setting the luminance value of N neighboring pixels to a value inversely proportional to the distance from the high luminance point C. The number N of neighboring pixels may be specified in advance, set by an operator, or automatically calculated based on a calculation formula given in advance from the total number of pixels of the target image. When the extended area overlaps with other high-luminance points when the brightness area is expanded, for example, the maximum value among the respective brightness values is set.

  FIG. 5 shows an example of expansion of the high luminance region, and a blackish circle and a whitish circle filled with dots represent high luminance points (hereinafter, the blackish circle is called a black circle and the whitish circle is called a white circle. It is assumed that the brightness value is higher than the white circle). The high luminance point to be subject to region expansion is determined by a method such as a pixel having a luminance value equal to or higher than a luminance threshold Sb set in advance, or the upper M luminance pixels in the target image. The luminance threshold value Sb and the luminance upper pixel number M are specified in advance or set by an operator, or automatically based on a calculation formula given in advance from the total pixel number or maximum luminance value of the target image. There are setting methods such as calculating automatically. For the teacher image 1 (FIG. 5A) and the teacher image 2 (FIG. 5B) for the target A, N pixels in the vicinity of each high-intensity point are newly set as high-intensity regions (FIG. 5C ), FIG. 5 (d)), and outputs them as expanded teacher image data E. A rectangular area centered on each high-luminance point represents the expanded high-luminance area.

  Subsequently, the type determination filter generation unit 5 generates a type determination filter F for the expanded teacher image data E. Specifically, each expanded teacher image data E is two-dimensionally Fourier transformed and vectorized by raster scanning to form a teacher image vector, and an inner product matrix between the teacher image vectors is obtained. The weight value of each teacher image vector is obtained so that the product of the inner product matrix and the vector composed of the weight values of each teacher image vector becomes a vector of all elements 1. Using the calculated weight value of each teacher image vector, the sum of the weighted teacher image vectors is obtained and used as a filter vector. Through the above procedure, the filter vector (that is, the type determination filter F) is calculated so that the correlation between the filter vector and all the teacher images is all 1. FIG. 5E shows a type determination filter F for the target A.

  In the above example, the plurality of post-expansion teacher image data E is integrated to generate the type determination filter F, but each post-extension teacher image data E may be used as an independent type determination filter as it is.

  Finally, the image matching unit 6 reads the observation image data that is the type determination target from the observation image data file 2 and performs type determination by pattern matching between the observation image data T and the type determination filter F. In pattern matching, the correlation coefficient is calculated while shifting the observed image data T and the type determination filter F one pixel at a time, and the correlation coefficient at the position where the correlation coefficient is the highest is obtained. If it is equal to or higher than the preset type determination threshold value Sr, it is determined as the corresponding model (type) and is output as the type determination result R, and the type determination result R is stored in the type determination result file 7.

  The type determination threshold value Sr is, for example, the degree of conformity of the image data that is not subject to filter calculation to the filter among the types of known image data obtained in advance when the type determination filter F is generated. For example, it is determined from the average correct answer rate or the like, and is set to a value having the highest degree of matching.

  As described above, in the target class identification device according to the first embodiment, the teacher image extension unit 4 extends the high-intensity region of the teacher image data, so that it is discrete due to the influence of the discrete distribution of the specification parameters. Interpolated high-intensity points are generated, so that it is possible to generate a general-purpose type determination filter that does not depend on the distribution state of the specification data. In addition, by setting the brightness value according to the distance from the high-brightness point when expanding the high-brightness region, the weighted high-brightness region can be expanded around the distribution point of the teacher image. The occurrence of false detection due to the divergence of. Furthermore, even when the number of teacher images for each item obtained in advance is limited, filter generation is possible by integration between teacher images with different items, so a filter with a high degree of fitness can be obtained.

Embodiment 2. FIG.
Since the configuration of the target class identification device according to the present embodiment is basically the same as the configuration of FIG. 1 described in the first embodiment, reference is made to FIG. 1 and description thereof is omitted here. . In the second embodiment, only the operation of the teacher image extension unit 4 is different from that in the first embodiment, and other operations are the same. Therefore, the following description will focus on operations different from those in the first embodiment. .

  In the above-described first embodiment (FIGS. 1 and 5), the teacher image expansion unit 4 expands the high luminance area to the neighboring pixels around the high luminance point. The distribution point of the high luminance point may be predicted in accordance with the fluctuation prediction, and the high luminance region may be expanded in the rotation direction. In the second embodiment, this case will be described.

  For example, the target spindle direction is considered as a specification parameter for which fluctuation is predicted. As a method of interpolating the target spindle direction distributed discretely, the locus of a high-luminance point by the target rotational movement is used. Specifically, in each teacher image, the rotation center point is determined from the positions of the M luminance points with the highest luminance (for example, the center point of the convex hull consisting of the M luminance points with the highest luminance, or the luminance upper point). The center of gravity of M high-luminance points, etc.), and each high-luminance point are rotationally moved by an angle ± θ, and the area defined by the rotational movement locus of the high-luminance points is expanded as a high-luminance area. Here, θ is designated in advance, set by an operator, or the direction of the principal axis is known or can be estimated, using the principal direction information (average angle difference, etc.) of the teacher image data in advance. A setting method such as automatic calculation using the calculation formula given in (5) is conceivable.

  FIG. 6 shows an example of expanding the high luminance region in the rotation direction. In FIG. 6, blackish circles and whitish circles filled with dots represent high luminance points (hereinafter, blackish circles are referred to as black circles and whitish circles are referred to as white circles. Black circles have higher luminance values than white circles). Further, it is assumed that the teacher images 1 and 2 in FIG. 6 are pre-processed by the image pre-processing unit 3. The operations up to the preprocessing in the image preprocessing unit 3 are the same as those in the first embodiment. The teacher image extension unit 4 performs the preprocessing by the image preprocessing unit 3 on the teacher image 1 (FIG. 6A) and the teacher image 2 (FIG. 6B) having different target main axis directions with respect to the target A. A center point (*) of rotation is obtained from the positions of the M high-intensity points of each brightness, and the high-intensity points are rotationally moved by an angle ± θ around the center point (FIG. 6 ( c), FIG. 6 (d)), the area defined by the rotational movement locus of each high-luminance point is expanded as a high-luminance area, and the expanded teacher image data E is generated. 6 (c) and 6 (d) is a region defined by the rotational movement locus of each high luminance point, that is, the high luminance after expansion in the present embodiment. It is an area.

  Next, the type determination filter generation unit 5 integrates the expanded teacher image data E shown in FIGS. 6C and 6D to generate the type determination filter F of the target A (FIG. 6E). . The filter generation procedure may be performed in the same manner as in the first embodiment. Note that the type determination filter F does not necessarily need to integrate a plurality of post-expansion teacher image data. Like the first embodiment, each type of post-extension teacher image data (FIGS. 6C and 6D) is used as it is. The determination filter F may be used. The image matching unit 6 performs type determination using the type determination filter F and stores the type determination result R in the type determination file 7. The operations after the type determination filter generation unit 5 are basically the same as those in the first embodiment.

  Further, the luminance value of the high luminance area is also equal to the luminance value of the high luminance point, or a value that is inversely proportional to the distance from the high luminance point, with the luminance value of the high luminance point being the maximum value, as in the first embodiment. May be set.

  As described above, according to the second embodiment of the present invention, the same effect as in the first embodiment can be obtained. In the second embodiment, the teacher image expansion unit 4 predicts the distribution points of the high luminance points in accordance with the fluctuation prediction of the observation specifications, and extends the high luminance region in the rotation direction. The extended range of the high luminance region is limited, and the probability of occurrence of erroneous detection can be suppressed.

Embodiment 3 FIG.
Since the configuration of the target class identification device according to the present embodiment is basically the same as the configuration of FIG. 1 described in the first and second embodiments, reference is made to FIG. Omitted. In the third embodiment, only the operation of the teacher image extension unit 4 is different from that in the first embodiment, and the other operations are the same. Therefore, the following description will focus on operations different from those in the first embodiment. .

  In the above-described Embodiments 1 and 2 (FIGS. 1 and 5 to 6), the teacher image extension unit 4 extends a high-luminance region to neighboring pixels around a high-luminance point, or high luminance in the rotation direction. Although the area has been expanded, it is not a relatively uniform movement of high-intensity points such as rotational movement, but the variation in luminance value between individual high-intensity points or the movement direction is the same, but the movement distance is high. Since there are situations such as different brightness points, the high brightness region may be extended in a specific direction in consideration of the movement distances. This case will be described in the third embodiment.

  For example, the incident angle of the transmission radio wave defined by the orbit of an observation radar mounted on a satellite or the like is considered as a specification parameter whose fluctuation is predicted. As the incident angle varies, a phenomenon called “falling” occurs, and the luminance value and distribution position of each high luminance point also vary. In this case, the luminance position of the high luminance point moves in the direction of the line of sight from the radar, and the moving distance of the high luminance point is the height direction position of the reflection point of the radio wave (that is, the reflection point on the target). It is known that it depends on the height from the ground). Therefore, the high brightness point of the teacher image is translated in the direction of the line of sight from the radar, and the area defined by the parallel movement locus of the high brightness point is expanded as a high brightness area.

  FIG. 7 shows an example of expanding the high luminance region in the line-of-sight direction. Here, it is assumed that two target teacher images have a relatively small incident angle difference. In FIG. 7, blackish circles and whitish circles filled with dots represent high luminance points (hereinafter, blackish circles are referred to as black circles and whitish circles are referred to as white circles. Black circles have higher luminance values than white circles). Further, it is assumed that the teacher images 1 and 2 in FIG. 7 are pre-processed by the image pre-processing unit 3. The operations up to the preprocessing in the image preprocessing unit 3 are the same as those in the first embodiment. The teacher image expansion unit 4 targets the teacher image 1 (FIG. 7 (a)) and the teacher image 2 (FIG. 7 (b)), which are pre-processed by the image preprocessing unit 3 and have different incident angles with respect to the target B. For each high-luminance point, a parallel movement of a distance ± α is performed in the line-of-sight direction to obtain a substantially rectangular belt-like region defined by the parallel movement locus of the high-luminance point (FIGS. 7C and 7D). ) To generate post-expansion teacher image data E in which the post-expansion high-intensity region is used. 7C and 7D, a band-like region including each high-luminance point in the center represents the high-luminance region after expansion in the present embodiment. Here, the movement distance α of each high-intensity point is based on, for example, the distribution state data of the reflection point height and the collapse distance obtained in advance if the height of the reflection point can be predicted to some extent in advance. , Predict and set the travel distance. If the height of the reflection point cannot be predicted, for example, the median value or the maximum value is set using the collapse distance range predicted from the target overall shape as an index.

  Next, the type determination filter generation unit 5 generates a type determination filter F for the target B (FIG. 7E) from the expanded teacher image data E. The filter generation procedure may be performed in the same manner as in the first embodiment. As in the first embodiment, the type determination filter F may use each expanded teacher image data as it is as the type determination filter F without generating a plurality of expanded teacher image data E by integrating them. The image matching unit 6 performs type determination using the type determination filter F and stores the type determination result R in the type determination file 7. The operations after the type determination filter generation unit 5 are basically the same as those in the first embodiment.

  Note that the example of FIG. 7 is intended for a case where the difference in incident angle is relatively small, but the incident angle of the teacher image as in the teacher image 3 and the teacher image 4 shown in FIGS. 8 (a) and 8 (b). When the difference opening is relatively large, divergence of the high-luminance region may occur due to filter generation for a plurality of post-expansion teacher image data.

  Here, considering that the incident angle is the information given at the time of observation as the radar observation specifications, a method of creating an independent filter without generating a filter between teacher images with a large difference in incident angle is also considered. It is done. That is, incident angle integration information that is information indicating an incident angle integration range at the time of filter generation is defined in advance, incident angles are grouped based on the incident angle integration information, and a filter is generated for each incident angle group (FIG. 8). (C) and FIG. 8 (d)) may be used.

  As for the luminance value of the high luminance region, the luminance of the high luminance point equal to or higher than the preset luminance normalization threshold Sn is considered to be a uniform value (in consideration of the luminance distribution variation due to the difference in incident angle). For example, a method of normalizing to Sn) (making it a luminance normalized value) and setting the luminance value in the extended high luminance region to the luminance normalized value can be considered. This makes it possible to generate a filter that is not affected by fluctuations in the luminance distribution.

  As described above, according to the third embodiment of the present invention, the same effect as in the first embodiment can be obtained. In the third embodiment, the teacher image extension unit 4 extends the high luminance region in a specific direction in consideration of the moving distance, thereby limiting the extension range of the high luminance region, and the occurrence probability of erroneous detection. Can be suppressed.

  In the third embodiment, the teacher image data is grouped for each incident angle, and the type determination filter is generated for each group. Therefore, when the incident angle difference is large, the brightness of each incident angle group is determined. Filter generation according to the distribution situation is possible. Here, the incident angle is given as an example of the specific observation parameter value, but the present invention is not limited to this, and grouping may be performed for each other observation parameter value. In the first and second embodiments, grouping may be performed, and in this case, the same effect can be obtained.

  In the third embodiment, the luminance value at the time of expansion of the high luminance region is normalized in consideration of the luminance distribution variation accompanying the variation in the incident angle, so that the filter does not depend on the luminance distribution variation. Generation is possible. In the description of the third embodiment, the incident angle is given as an example of the specific observation specification parameter value. However, the present invention is not limited to this, and the luminance distribution change accompanying the change of other specific observation specification parameter value is not limited. In order to compensate, the luminance value at the time of expanding the high luminance region may be normalized. Also in the first and second embodiments, normalization may be performed, and in this case, the same effect can be obtained.

  In addition, this invention is not limited to said each Embodiment 1-3, It cannot be overemphasized that all the possible combinations of each Embodiment 1-3 are included.

  1 teacher image data file, 2 observation image data file, 3 image preprocessing unit, 4 teacher image expansion unit, 5 type determination filter generation unit, 6 image collation unit, 7 type determination result file, F type determination filter, K after extension Teacher image data, L teacher image data, P pre-processed image data, R type determination result, T observation image data.

Claims (6)

A target class identification device for determining the type of observation target for observation image data of an observation target,
A teacher image data file for storing a supervised image data group composed of supervised image data of observation targets obtained by observation under a plurality of observation conditions, and its observation data;
An observation image data file for storing the observation image data to be subjected to the type determination and the observation specification data;
An image preprocessing unit that performs preprocessing before filter generation such as alignment on the teacher image data, and generates preprocessed teacher image data after preprocessing;
A teacher image extension unit that generates a post-expansion teacher image data by extending a high-luminance region to compensate for the discrete distribution of the observation data of the pre-processing teacher image data. ,
A type determination filter generation unit that generates a type determination filter from the expanded teacher image data;
Collating the observation image data with the type determination filter, and an image verification unit for determining the type of the observation target;
And a type determination result file for storing a type determination result of the observation target.   The teacher image extension unit obtains a center point of rotation of the high brightness point based on the position of the high brightness point in the pre-processed teacher image data, and centers the center point for each high brightness point. The target class according to claim 1, wherein the high brightness area is expanded by performing a rotational movement of a predetermined angle and setting the area defined by the rotational movement locus of the high brightness point as the expanded high brightness area. Identification device.   The teacher image extension unit translates each high-intensity point in the pre-processed teacher image data by a predetermined distance in a specific direction based on the viewing direction at the time of observation, and translates the high-intensity point The target class identification device according to claim 1, wherein the high-luminance region is expanded using the region defined by the locus as the high-luminance region after expansion.   The teacher image extending unit extends the high brightness region by changing a brightness value of the high brightness region in accordance with a distance from each high brightness point in the pre-processed teacher image data. Item 1. A target class identification device according to Item 1.   2. The target class identification according to claim 1, wherein the teacher image extension unit groups the teacher image data in accordance with a specific observation specification parameter value and generates the type determination filter for each group. apparatus.   2. The target class according to claim 1, wherein the teacher image extension unit normalizes the luminance value of each high luminance point in order to compensate for luminance distribution variation accompanying variation of a specific observation parameter. Identification device.

Images