JP5076972B2 - Target detection system - Google Patents

Description

  The present invention relates to a technique for detecting a target pattern from an image, and more particularly to a pattern detection technique suitable for a panchromatic satellite image.

  Land observation and earth observation satellites such as QuickBird (DigitalGlobe, USA), ALOS “Daichi” (Japan AXA), and TERRA / ASTER (Japan ERSDAC) orbit around the Earth in a polar orbit at an altitude of about 450 km. It is equipped with sensors and synthetic aperture radar (SAR), continuously scanning the ground, and transmitting image data to the earth station on a regular basis. Three types of images are mainly used, and the optical sensor's multi-spectral mode (color photo) can be photographed at night in a 4-band configuration that adds the near-infrared to the RGB band of visible light. It is effective for collecting disasters and other situations. The SAR image is a radar image generated by radio waves and transmits even when it is cloudy. Therefore, the image can be collected even when the weather is bad, and a ship in operation can be measured using the Doppler effect (see Patent Document 1). . Among them, the panchromatic mode (monochrome photo) of the optical sensor has a resolution of less than 1 m (the world's highest resolution QuickBird 0.6 m) that can identify passenger cars on the ground, and is useful for monitoring traffic conditions. With regard to traffic monitoring, aerial photography with a surveillance camera, helicopter, or aircraft installed on the ground provides clearer images, but it is difficult to observe a wide area at the same time, and it is hidden behind buildings. In recent years, satellite images have attracted attention due to the problem that a blind spot always occurs.

In addition, a target detection method using template matching that specializes in aircraft and ships (see Patent Document 2), a template is prepared in advance, and pixel blocks of the same size as the template are extracted while sequentially scanning from the first pixel of the image data Also, a method of matching with a template (see Patent Document 3) has been proposed.
Japanese Patent No. 3863014 Japanese Patent No. 3978979 JP 2001-67470 A

  However, in the method of Patent Document 2, even if the same individual target is detected twice, each of them is presented as a detection result, so that it is ultimately necessary for a person to judge There is. In addition, the technique disclosed in Patent Document 3 has a problem that it cannot be applied to detection of a target because it only determines whether a target is acceptable or not.

  In particular, when detecting a target from a satellite image, a target of the same individual may be detected as a separate target at a position shifted by several pixels from the detection position of a certain target. In some cases, a separate target is actually present at a position deviated, and it is difficult to discriminate the target. For example, when a ship is anchored adjacent to a quay, the same ship may be detected twice or an adjacent ship may not be detected.

  Therefore, an object of the present invention is to provide a target detection system that can accurately detect a target of each individual without double detection without omission.

In order to solve the above-described problem, in the first aspect of the present invention, a template storage unit that stores a template image is a system that detects a target from a search target image that is a search target by comparison with the template image. And extracting a pixel block of the same size as each of the template images from the search target image while sequentially moving the extraction position from the start pixel position to the final pixel position with respect to the search target image, Sequential similarity determination means for calculating similarity evaluation value of a pixel block and comparing the similarity evaluation value with a predetermined criterion value, and similarity by the sequential similarity determination means If it is determined that there is a similarity, each template position is determined for each pixel position within a predetermined range from the pixel position determined to be similar. A pixel block having the same size as the image image is extracted from the search target image, a similarity evaluation value between the template image and the pixel block is calculated, and a pixel position at which an optimal similarity evaluation value is obtained is detected as a target. the local similarity determining means for determining the position, the relates target object detected by the local similarity determination means, the search result output pixel position location to identify the location of the pixel block in the previous SL on the search target image at that time A detection target that erases the detected target by setting each pixel in the pixel block on the search target image corresponding to the pixel position that is the output target and the detected target detection position to a predetermined value possess and objects erasing means, and said sequential similarity determination means, the local similarity determination means, said pixel block and the corresponding pixels included in said template image Calculating a gradation fluctuation histogram for each pixel, the Euclidean distance of the gradation variation histograms each other both to provide a target detection system is calculated as the similarity evaluation value.

  According to the first aspect of the present invention, when a target is detected from the search target image by comparing the template image including the target and the pixel block in the search target image, the target is detected at the optimum position, and the detection is performed. Since the target is deleted from the search target image, it is possible to accurately detect the target of each individual without omission without double detection.

  In the second aspect of the present invention, when the sequential similarity determination means in the first aspect of the present invention calculates a plurality of the similarity evaluation values and it is determined that there is no similarity with one similarity evaluation value, the other It is characterized in that it is determined that there is no similarity without making a determination on the similarity evaluation value. According to the second aspect of the present invention, when a plurality of similarity evaluation values are calculated and it is determined that there is no similarity based on one similarity evaluation value, it is determined that there is no similarity as a whole. Since the determination regarding the similarity evaluation value is not performed, the overall processing speed is increased.

  In the third aspect of the present invention, a polygonal effective image region is set in the template image in the first or second aspect of the present invention, and the sequential similarity determination means and the local similarity determination means are The similarity evaluation value is calculated based on a pixel group included in the effective image area in the template image and a pixel group corresponding to the inside of the effective image area in the pixel block. Features. According to the third aspect of the present invention, the peripheral pixels of the effective image area included in the template image are excluded from the calculation of the similarity evaluation value in the first or second aspect of the present invention. Calculation is possible.

  According to the present invention, when a target is detected from a search target image by comparing a template image including the target and a pixel block in the search target image, the target of each individual is not detected, and there is no omission. It becomes possible to detect accurately.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
(1. System configuration)
First, the configuration of the target object detection system according to the present invention will be described. FIG. 1 is a configuration diagram of a target object detection system according to the present invention. In FIG. 1, 10 is a search target image storage unit, 20 is a template preparation unit, 30 is a template storage unit, 40 is a target search unit, and 50 is a search result storage unit.

  The search target image storage unit 10 stores a search target image that is an image to be searched and a sample image that stores a sample of the target. If there are many targets to be searched on the search target image, one of them can be used as a sample. In this case, the search target image also serves as a sample image. In the present embodiment, a satellite image that is an image taken by an artificial satellite is stored as a search target image. The template preparation unit 20 interactively sets a sample image to prepare a template image, and includes a template creation unit 21, a determination reference value setting unit 22, and a template update unit 23. The template storage unit 30 stores the template image prepared by the template preparation unit 20. The target searching unit 40 searches for a target by comparing a predetermined region in the search target image with the template image. The target searching unit 40 sequentially searches for a similarity determination unit 41, a local similarity determination unit 42, and a search result output. Means 43, detected object erasing means 44, and second template updating means 45 are provided. The search result storage unit 50 stores the results searched by the target object search unit 40. The target object detection system shown in FIG. 1 is actually realized by incorporating a dedicated program into a general-purpose computer. Each storage unit is realized by a storage device such as a hard disk built in or connected to the computer.

(2. Preparation of template image)
Next, the processing operation of the target object detection system shown in FIG. 1 will be described. First, the search target image storage unit 10 stores a satellite image captured by an artificial satellite as a search target image. In the present embodiment, the satellite image is in the Geo-TIFF format compliant with the Adobe TIFF standard, and the image data with geographic information (such as latitude / longitude information) added thereto is adopted.

  When the target detection system is activated, first, the template preparation unit 20 converts the search target image stored in the search target image storage unit 10 from the source image format to the work image format. This is done to increase the search efficiency. Specifically, it is converted into a work image format using a known histogram flattening technique. In the present embodiment, the work image is created in the RXY format. The template preparation unit 20 uses this search target image as a sample image.

  Subsequently, the template preparation unit 20 displays the search target image as the sample image converted into the work image format on the screen, and prompts the user to specify an image area for creating the template image. On the other hand, the user designates the contour of an image area having a polygonal shape (in this embodiment, a free rectangular shape including four vertices) in the search target image displayed on the screen. This is performed by designating four vertices of a square shape by operating the mouse. For example, when a ship is an object, the user finds a ship shown in a satellite image and designates an effective image area including the ship. Specifically, a square of a square shape is displayed as an initial state on the screen, and the user selects one of the vertices on the screen, designates the vertex by dragging the vertex to a desired position, and designates four vertices. The same operation is sequentially performed on the vertices. When the designation is made, the template creation means 21, as shown in FIG. 2, circumscribes the square shape (outside of FIG. 2) so as not to protrude even if the square is rotated 360 degrees outside the designated square. Square shape) at the same time. Accordingly, a double square is displayed on the screen in the initial state, and when the user transforms the inner square shape into a desired quadrangle, the outer square shape changes while maintaining the square.

  When the user completes the specification of a predetermined quadrangular effective image area in the displayed search target image, the template creation means 21 defines a square defined outside the effective image area specified from within the search target image. Trimming based on shape. As a result, a square-sized template image is obtained, and at the same time, information for masking the outer portion of the rectangular area defined by the user in the template image is added. Specifically, as shown in FIG. 2, each pixel value of the template image has an 8-bit value from 0 to 255, but the inside of the square area defined by the user has a pixel value from 1 to 255. The outer side has a pixel value of 0.

  Furthermore, the template preparation unit 20 also registers in the template storage unit 30 a polygon shape and a pattern obtained by rotating only the inside of the polygon shape at a predetermined angle. In the present embodiment, the angle is rotated in units of 45 degrees, and template images according to eight different inclination angles are created. FIG. 3 shows eight template images having different angles. At this time, since the outer square shape is defined so that it does not protrude even if the polygon specified by the user is rotated 360 degrees, the polygons of the eight types of angles are represented by template images of the same size. can do. In this embodiment, patterns with different angles are registered with the same template ID as the same template image. Therefore, the template images shown in FIG. 3 all have the same template ID, and each pattern is specified by its angle. In this embodiment, eight template IDs can be registered. In this case, since eight template images are registered at eight angles, 64 patterns are registered in total.

  When the template image is registered, the template preparation unit 20 prompts the user to specify an object to be matched on the search target image. When the user designates an object to be matched, the determination reference value is calculated and the template image is updated using the pixel values near the designated position. The calculation of the determination reference value and the update of the template image will be described with reference to the flowchart of FIG.

  First, the user looks at the contents of the search target image displayed on the screen, finds a target object that he wants to include in the search target, and designates a point on the target object (S101). FIG. 5A is a diagram illustrating a state of a search target image when an object is specified. When the user designates a point on the target object to be included in the search target image in the search target image, as shown in FIG. 5A, a polygonal shape indicating a square outer frame of the template image and an effective image area is obtained. Is displayed. For example, when a ship is to be detected from a satellite image, a point on one ship in the satellite image is designated.

  Subsequently, the template preparation unit 20 performs matching processing with the template image using the designated position as a base point (S102). The reason for this processing is that the point specified by the user on the screen is based on visual judgment and is ambiguous as shown in FIG. 5A. Therefore, the judgment reference value is calculated as it is based on this point. This is because there is a high possibility that an inappropriate search is performed. Therefore, in this process, the point designated by the user is corrected to an optimum position. Specifically, as shown in FIG. 5B, the base point (Xc, Yc) on the designated search target image and the center position of the template image are aligned, and from there, the template image is set to 1 in the x and y directions. Matching processing is performed while moving within a set range in pixel units. Details of this matching process will be described later. At the same position, matching processing is performed on eight types of template images with different angles, a template image having the highest similarity is selected, and the pixel position of the search target image is specified as the optimum position.

  Next, the determination reference value setting unit 22 sets a determination reference value so that the object on the search target image is determined to be compatible (S103). When the determination reference value is set, the template updating unit 23 combines the pixel block of the search target image at the optimum position and the template image, and updates the template image with the combined image (S104).

  Next, the details of the matching process in S102 will be described with reference to the flowchart of FIG. When performing the matching process, first, each variable that becomes the similarity evaluation value is initialized. Specifically, the minimum value of the luminance histogram difference, the minimum value of the gradation fluctuation histogram, the minimum value of the distance between the centroids, the minimum value of the moment of inertia difference, the minimum value of the gray gradation pixel difference, and the normalized phase relationship The six variables (similarity evaluation values) having the maximum number are initialized. In the present embodiment, all of these six variables take values from 0 to 1000, with respect to the normalized correlation coefficient, the initial value is the minimum value 0, and for the other five variables, the initial value is the maximum value 1000. To do.

  When the initialization of each variable is completed, the luminance histogram is collated (S201). Specifically, first, the size of the template image is set to S × S, and the gradation value at the pixel (x, y) (0 ≦ x ≦ S−1, 0 ≦ y ≦ S−1) of the template image is set to Vt ( x, y) (0 to 255), and the image gradation value corresponding to the template position (x, y) in the region starting from the position of the search target image (Xc + dx, Yc + dy) is V (x + Xc + dx, y + Yc + dy) (1 to 255). Within the range of S × S pixels, V (x, y) / 16 of the pixel (x, y) where Vt (x, y)> 0 corresponding to the pixel in the effective image area shown in FIG. 16-gradation luminance histogram H (v) (0 ≦ v ≦ 15) of the search target image based on the value and 16-gradation luminance histogram Ht (v of the template image based on the value of Vt (x, y) / 16 ) (0 ≦ v ≦ 15) is calculated. Further, a luminance histogram difference is calculated by executing processing according to the following [Equation 1].

[Formula 1]
(Luminance histogram difference) = 1000 · _{Σv = 0,15} | H (v) −Ht (v) | / _{Σv = 0,15} (H (v) + Ht (v))

  Subsequently, the gradation fluctuation histogram is collated (S202). Specifically, as in S201, the size of the template image is S × S, and Vt (x, y)> 0 corresponding to the pixels in the effective image area shown in FIG. 2 within the range of S × S pixels. 16 pixel gradation fluctuation histogram HD (v) (0 ≦ v ≦) of the search target image based on the value of V (x, y)% 16 (the remainder divided by 16) in the pixel (x, y) 15) and a gradation fluctuation histogram HDt (v) (0 ≦ v ≦ 15) of 16 gradations of the template image based on the value of Vt (x, y)% 16. HD (v) is a gradation fluctuation histogram based on the lower 4 bits of each pixel, and HDt (v) is an upper 4 bits of each pixel. Furthermore, the gradation fluctuation histogram difference, that is, the Euclidean distance between the gradation fluctuation histograms is calculated by executing processing according to the following [Equation 2].

[Formula 2]
(Gradation fluctuation histogram difference) = 1000 · _{Σv = 0,15} | HD (v) −HDt (v) | / _{Σv = 0,15} (HD (v) + HDt (v))

  Next, the distance between the centers of gravity is calculated (S203). Specifically, first, the size of the template image is set to S × S, and the gradation value at the pixel (x, y) (0 ≦ x ≦ S−1, 0 ≦ y ≦ S−1) of the template image is set to Vt ( x, y) (0 to 255), the image gradation value corresponding to the position (x, y) of the template image in the region starting from the position (Xc + dx, Yc + dy) of the search target image is represented by V ( x + Xc + dx, y + Yc + dy) (1 to 255). Within the range of S × S pixels, using the pixel (x, y) for which Vt (x, y)> 0 corresponding to the pixel in the effective image area shown in FIG. , Yg) and the center of gravity (Xgt, Ygt) of the template image are calculated by executing processing according to the following [Equation 3].

[Formula 3]
Xg = Σy = _{0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy) ・ x
/ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy)
Yg = Σy = _{0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy) ・ y
/ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy)
_{Xgt = Σ y = 0, S} -1 Σ x = 0, S-1 | Vt (x, y) ≠ 0 Vt (x, y) · x / Σ y = 0, S-1 Σ x = 0, S _{-1 | Vt (x, y) ≠ 0} Vt (x, y)
_{Ygt = Σ y = 0, S} -1 Σ x = 0, S-1 | Vt (x, y) ≠ 0 Vt (x, y) · y / Σ y = 0, S-1 Σ x = 0, S _{-1 | Vt (x, y) ≠ 0} Vt (x, y)

  Furthermore, the process according to the following [Equation 4] is executed to calculate the distance between the centers of gravity.

[Formula 4]
(Distance between centroids) = 1000 · {(Xg−Xgt) (Xg−Xgt) + (Yg−Ygt) (Yg−ygt)} ^1/2 / S

  Next, the inertia moment is collated (S204). Specifically, first, using the centroid (Xg, Yg) of the search target image obtained in S203 and the centroid (Xgt, Ygt) of the template image, within the range of S × S pixels, the validity shown in FIG. Using the pixel (x, y) with Vt (x, y)> 0 corresponding to the pixel in the image area, the inertia moment M of the search target image and the inertia moment Mt of the template image are expressed by the following [Equation 5]. Execute the process according to

[Formula 5]
M = Σy _{= 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy) ・ [(x-Xg) (x -Xg) + (y-Yg) (y-Yg)]
/ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy)
Mt = Σy _{= 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} Vt (x, y) ・ [(x-Xgt) (x-Xgt) + (y-Ygt ) (y-Ygt)]
/ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} Vt (x, y)

  Further, a process according to the following [Equation 6] is executed to calculate the moment of inertia difference.

[Formula 6]
(Inertia moment difference) = 1000 · | M−Mt | / (M + Mt)

  Steps S201 to S204 are processes that do not depend on the angle. Therefore, it is only necessary to perform once for each template image. Since the template image at any angle may be used in the processes of S201 to S204, a template image of 0 degree is typically used in the present embodiment. Since S205 and S206 from here are processing depending on the angle, the template image having the same ID is also processed for each angle. First, a gray gradation pixel difference is calculated (S205). Specifically, first, the size of the template image is set to S × S, and the gradation value at the pixel (x, y) (0 ≦ x ≦ S−1, 0 ≦ y ≦ S−1) of the template image is set to Vt ( x, y) (0 to 255), and the image gradation value corresponding to the template position (x, y) in the region starting from the position (Xc + dx, Yc + dy) of the search target image is represented by V (x + Xc + dx, y + Yc + dy) (1 to 255). Here, the following tone conversion is performed to raise the shadow side to the highlight side and drop the halftone portion to the shadow side. When Vt (x, y) ≥128, Vt '(x, y) = Vt (x, y) -128, when Vt (x, y) <128, Vt' (x, y) = 127-Vt (x, y) When V (x, y) ≥128, V '(x, y) = V (x, y) -128, when V (x, y) <128, V' (x, y ) = 127−V (x, y) Within the range of S × S pixels, C pixels (x, x) corresponding to the pixels in the effective image area shown in FIG. Using y), processing according to the following [Equation 7] is executed to calculate a gray gradation pixel difference which is a pixel difference between the search target image and the template image.

[Formula 7]
(Gray gradation pixel difference) = 1000 · Σ _{y = 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} | V '(x + Xc + dx, y + Yc + dy ) −Vt '(x, y) |
/ (V '(x + Xc + dx, y + Yc + dy) + Vt' (x, y))

  Next, a normalized correlation coefficient is calculated (S206). Specifically, first, the size of the template image is set to S × S, and the gradation value at the pixel (x, y) (0 ≦ x ≦ S−1, 0 ≦ y ≦ S−1) of the template image is set to Vt ( x, y) (0 to 255), and the image gradation value corresponding to the template position (x, y) in the region starting from the position (Xc + dx, Yc + dy) of the search target image is represented by V (x + Xc + dx, y + Yc + dy) (1 to 255). Within the range of S × S pixels, C pixels (x, y) satisfying Vt (x, y)> 0 corresponding to the pixels in the effective image area shown in FIG. The average gradation value Va and the average gradation value Vat of the template image are calculated by executing processing according to the following [Equation 8].

[Formula 8]
Va = Σy _{= 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} V (x + Xc + dx, y + Yc + dy) / C
Vat = Σy _{= 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} Vt (x, y) / C
C = Σy _{= 0, S-1} Σx _{= 0, S-1 | Vt (x, y) ≠ 0} 1

  Then, within a range of S × S pixels, a pixel (x, y) that satisfies Vt (x, y)> 0 corresponding to the pixel in the effective image area shown in FIG. Based on this, the normalized correlation coefficient is calculated by executing processing according to the following [Equation 9]. However, when the normalized correlation coefficient is a negative value (negative correlation), it is treated as 0 (no correlation).

[Formula 9]
(Normalized correlation coefficient) = 1000 ・ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} (V (x + Xc + dx, y + Yc + dy) -Va) ・ (Vt (x, y) -Vat)
/ [Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} (V (x + Xc + dx, y + Yc + dy) -Va) ²・ Σ _{y = 0, S-1} Σ _{x = 0, S-1 | Vt (x, y) ≠ 0} (Vt (x, y) -Vat) ² ] ^1/2

  Subsequently, each variable is updated (S207). Specifically, the luminance histogram difference calculated in S201 is smaller than the minimum value of the currently set luminance histogram difference, and the normalized correlation coefficient calculated in S206 is the normalized correlation coefficient currently set. If the value is larger than the maximum value, the minimum value of the brightness histogram difference, the minimum value of the gradation fluctuation histogram difference, the minimum value of the distance between the centroids, the minimum value of the moment of inertia difference, the minimum value of the gray gradation pixel difference, and the normalization phase Replace the six variables with the maximum number of relations with the values calculated this time. Then, the position, template ID, and angle values are updated with the position, template ID, and angle values at this time.

  When the update of each variable is completed, it is determined whether or not the processing of all angle patterns has been completed for one template image (S208). If the processing has not been completed for all the angle patterns, the process returns to S205 to calculate the gray gradation pixel difference depending on the angle and the normalized correlation coefficient. If the process is completed for all angle patterns, the process proceeds to S209. Therefore, in the present embodiment, the processing of S205 to S207 is repeated for eight angles for one template image.

  When processing is completed for all angles of one template image, it is determined whether or not processing has been completed for all template images (S209). If the processing has not been completed for all template images, the processing of S201 to S208 is repeated. If the processing is completed for all template images, the process proceeds to S210. Therefore, in the present embodiment, the processing of S201 to S208 is repeated for eight template images for one pixel position.

  When processing is completed for all template images, it is determined whether processing has been completed for all pixel positions (S210). All pixel positions mean the positions of pixels included in a predetermined pixel range in the x-axis and y-axis directions from the designated base point (Xc, Yc). As the predetermined pixel range, about 5% of each of the x-axis and y-axis directions of the template image is set. For example, when the template image is 100 × 100 pixels, the processing shown in FIG. 5 is repeated for a total of 25 × 5 × 5 pixel positions. If the processing has not been completed for all pixel positions, the processing of S201 to S209 is repeated. When the process is completed for all pixel positions, each variable is output and the matching process is terminated. Each variable is output for each template image.

  Next, details of setting the determination reference value in S103 will be described. The determination reference value is set by multiplying each variable output in the matching process by a predetermined value by the determination reference value setting means 22. In the present embodiment, for the normalized correlation coefficient, a value obtained by multiplying the maximum value of the output normalized correlation coefficient by 0.9 is used as a determination reference value, and for the other five variables, the output value A value obtained by multiplying 1.1 by 1.1 is used as a criterion value. Specifically, the calculation is performed by executing processing according to the following [Equation 10].

[Formula 10]
(Determination reference value of luminance histogram difference) = (Minimum value of luminance histogram difference) × 1.1
(Judgment reference value of gradation fluctuation histogram difference) = (minimum value of gradation fluctuation histogram difference) × 1.1
(Judgment reference value of distance between centroids) = (minimum value of distance between centroids) × 1.1
(Judgment reference value of moment of inertia difference) = (minimum value of moment of inertia difference) × 1.1
(Determination reference value of gray gradation pixel difference) = (Minimum value of gray gradation pixel difference) × 1.1
(Criteria value of normalized correlation coefficient) = (maximum value of normalized correlation coefficient) × 0.9

  Next, details of the template update in S104 will be described. In S104, the template update unit 23 updates each template image. Specifically, the image of each angle of each template image is updated using the pixel block on the search target image at the optimum position of each template image. First, among the template images, the one with the optimum angle and the pixel block on the search target image are synthesized. Since the mask area is not synthesized in the template image, the composition is performed only within the effective image area in the template image.

  As a specific method of synthesis, an integer pseudo-random number is sequentially generated for each pixel in the effective image area, and when a predetermined number is obtained, it is replaced with the value of the corresponding pixel in the pixel block on the search target image. For example, if an odd number appears, if it is set to be replaced with a corresponding pixel of a pixel block on the search target image, a pixel in the effective image area is replaced with a pixel of the search target image with about half probability. FIG. 7 is a diagram illustrating an example of updating the template image by combining the search target image and the template image. In the example of FIG. 7, the image size is simplified to 5 × 5 pixels for convenience of explanation. FIG. 7B shows a template image, T indicates a pixel in the effective image area (polygonal area), and a blank indicates a pixel in the mask area. FIG. 7A shows a pixel block on the search target image determined to be the optimum position with respect to the template image, and W indicates a pixel on the search target image. As a result of performing the above-described combining process on these two images, as shown in FIG. 7C, about half of the pixels T in the effective image area of the template image are replaced with the pixels W of the search target image. A composite image is obtained. This composite image is registered as an updated template image. Note that the seeds are changed each time so that the random number generation method becomes a different series each time processing is performed.

  When the update process is completed for the optimum angle pattern, the other angle patterns are also updated. For other angle patterns, obtain the difference between the angle and the optimum angle, rotate the pixel block on the search target image by that angle, and rotate the pixel unit inside the effective image area to rotate the pixel block on the search target image. Replace with the corresponding pixel. In this way, the template image is updated for all angles. All template images are updated by performing the same process on all template images prepared in advance. FIG. 8 is a diagram showing how the template image is updated. As shown in FIG. 8, images at all angles are updated using the pixel block at the optimum position on the search target image.

(3. Search process)
As described above, when the preparation of the template image is completed, the target object search unit 40 searches for the target object using the prepared template image. The target searching process will be described with reference to the flowchart of FIG. When the user instructs the target searching process, the sequential similarity determining unit 41 of the target searching unit 40 sequentially performs the matching process (S301). Specifically, pattern matching processing is sequentially executed from the start pixel position in the search target image to find a position for determining similarity.

  FIG. 10 is a diagram showing an outline of the sequential matching process. In the sequential matching process, first, a square pixel block up to (S-1, S-1) from the start pixel position (0, 0) in the search target image is compared with the template image, A square area starting at (1, 0), a square area starting at (2, 0), and so on. Then, it moves to the Y direction scan range by one pixel, and the entire search target image is sequentially compared with the template image such as a square area starting from (0, 1), a square area starting from (1, 1), and so on. Make a comparison.

  In the matching process at each pixel position, first, angle-independent template matching for matching with one of the template images with the same ID and angle-dependent template for matching with all eight of the template images with the same ID. It is divided into matching. If it is determined that the angle-independent template matching is incompatible, the position shifts to the next pixel position without performing the angle-dependent template matching.

  Details of the sequential matching process when the pixel position is specified are shown in the flowchart of FIG. First, as in the matching process in S102, each variable that becomes a similarity evaluation value is initialized. Specifically, the minimum value of the luminance histogram difference, the minimum value of the gradation fluctuation histogram, the minimum value of the distance between the centroids, the minimum value of the moment of inertia difference, the minimum value of the gray gradation pixel difference, and the normalized phase relationship Initialize the six variables with the maximum number. All of the six similarity evaluation values used in the search process also take a value of 0 to 1000. For the normalized correlation coefficient, the initial value is set to the minimum value 0, and for the other five variables, the initial value is set to the maximum value 1000. And Subsequently, the brightness histogram is collated (S401). Specifically, as in S201 above, the process according to [Formula 1] is executed to calculate the luminance histogram difference. Then, the luminance histogram difference is compared with the determination reference value set in S103. Nonconforming means that there is no similarity. If it does not match, it is determined that the currently used template image does not match, and it is determined whether or not processing with all template images has been completed (S409). If processing with all template images has not been completed, the template image is changed to another template image (S410), and the process returns to S401.

  If the luminance histogram difference is smaller than the determination reference value in S401, the gradation fluctuation histogram is collated (S402). Specifically, similar to S202 above, the gradation fluctuation histogram difference is calculated by executing the processing according to [Formula 2]. Then, the gradation fluctuation histogram difference is compared with the determination reference value set in S103. Processing in the case of nonconformity is the same as that in the case of nonconformity in S401.

  If the gradation fluctuation histogram difference is smaller than the determination reference value in S402, the distance between the centers of gravity is calculated (S403). Specifically, similarly to S203, the processing according to [Formula 3] and [Formula 4] is executed to calculate the distance between the centers of gravity. Then, the distance between the centers of gravity and the determination reference value set in S103 are compared. Processing in the case of nonconformity is the same as that in the case of nonconformity in S401.

  If the distance between the centers of gravity is smaller than the determination reference value in S403, the inertia moment is collated (S404). Specifically, similar to S204, the processing according to [Formula 5] and [Formula 6] is executed to calculate the moment of inertia difference. Then, this moment of inertia difference is compared with the determination reference value set in S103. Processing in the case of nonconformity is the same as that in the case of nonconformity in S401. Since the processing of S401 to S404 is processing that does not depend on the angle, if the template image is not matched even once after matching with the template image, the process proceeds to processing with another template image.

  If the moment of inertia difference is smaller than the determination reference value in S404, the gray gradation pixel difference is calculated (S405). Specifically, similarly to S205 described above, the processing according to [Formula 7] is executed to calculate the gray gradation pixel difference. In the gray tone pixel difference calculation process, unlike the processes in S401 to S404, the same template image is processed for all angles. The gray gradation pixel difference is updated to the minimum, and the angle at that time is recorded. When the calculation of the gray gradation pixel difference is completed for all angles (S406), the minimum gray gradation pixel difference is compared with the determination reference value set in S103 to obtain the gray gradation pixel difference. If is more than the criterion value, it is not conforming. Processing in the case of nonconformity is the same as that in the case of nonconformity in S401.

  When the gray gradation pixel difference is smaller than the determination reference value, a normalized correlation coefficient is calculated (S407). Specifically, the normalized correlation coefficient is calculated by executing processing according to [Equation 8] and [Equation 9] as in S206. In the normalized correlation coefficient calculation process, similar to the gray gradation pixel difference calculation process in S405, the same template image is processed for all angles. The normalized correlation coefficient is updated to the maximum value, and the angle at that time is recorded. Then, the processing when the normalized correlation coefficient has been calculated for all the angles is the same as that in the case of nonconformity in S401.

  When the processing is completed for all template images, the normalized correlation coefficient is collated (S411). Specifically, the maximum normalized correlation coefficient is compared with the determination reference value set in S103, and if the normalized correlation coefficient is larger than the determination reference value, the result is relevant. If the normalized correlation coefficient is matched, it is determined whether there is a template that matches all items, that is, all six similarity evaluation values (S412). If there is a template in which all items are compatible, it is determined whether or not the angle of the template in which all items are compatible matches that of the template (S413). That is, the template and the angle that give the minimum value in the calculation processing of the gray gradation pixel difference in S405 and S406 and the template and the angle that give the maximum value in the calculation processing of the normalized correlation coefficient in S407 and S408 respectively match. When it is determined in S401, S402, S403, S404, S406, and S411 that the matching templates are all matched, matching ends, and each variable, that is, template ID, pixel position (Xc, Yc), angle, and six similarities. The sex evaluation value is output. In any case, in the case of non-conformity, the non-conformance ends, and matching processing is sequentially performed on the next pixel position.

  In the sequential matching processing shown in FIG. 11, the similarity determination unit 41 sequentially compares the luminance histogram in S401, the gradation fluctuation histogram in S402, the calculation of the distance between the centers of gravity in S403, and the inertia moment verification in S404. Six suitability determination processes are performed, that is, the calculation of the gray tone pixel difference in S405 and S406, and the calculation of the normalized correlation coefficient in S407 and S408. These suitability determination processes are performed in order of increasing processing load. It is set to be done. That is, in the present embodiment, the comparison of luminance histograms has the smallest processing load, and the calculation of the normalized correlation coefficient has the largest processing load. This is because, when it is determined that the non-conformity is determined in one process, it is more efficient to perform the process starting from the one with the smallest processing load if it is determined as dissimilar as a whole. In particular, since the calculation of the gray gradation pixel difference and the calculation of the normalized correlation coefficient are performed for each angle, the comparison of the brightness histogram, the comparison of the gradation fluctuation histogram, and the distance between the centroids performed only for one angle are performed. Compared to calculation and moment of inertia verification, the processing load is significantly increased.

  If the sequential matching process in S301 is completed, the local similarity determination unit 42 performs the local matching process (S302). The sequential matching process in S301 searches for a rough position of the target in the entire search target image, whereas the local matching process in S302 searches for a detailed position of the target. Specifically, template matching is performed by moving a predetermined minute width (ΔX, ΔY) in the pixel scan progression direction (rightward, downward) starting from the position (Xc, Yc) determined to be optimal by sequential matching processing. And the six similarity evaluation values are updated to the minimum value and the maximum value, and the template ID, pixel position (Xm, Ym), and angle Am at that time are output. The details of the matching process are the same as the matching process in S102 of FIG. 4, specifically the process shown in FIG. 6, unlike the sequential matching process shown in FIG. The sequential matching process only has to determine whether or not the matching is performed at a certain position (Xc, Yc), but the local matching process searches for a matching position under the best condition within a minute range in which the matching is determined. This is necessary.

  When the local matching process is completed, the search result output unit 43 of the target object search unit 40 outputs the search result (S303). Specifically, the template ID, position (Xm, Ym), angle Am, and six similarity evaluation values are output as a list.

  Subsequently, the detection target erasing unit 44 erases the detection area (S304). Specifically, in the search target image, a portion corresponding to the inside of the effective image area shown in FIG. 2 of the template image at the optimum position is deleted. That is, all the pixel values on the search target image located inside the effective image area are set to zero. FIG. 12 is a diagram showing the optimum position of the search result on the search target image and the erased area. As a result of the search, when the position shown in FIG. 12A is the optimum position of the effective image area of the template image with respect to the target on the search target image, as shown in FIG. A portion corresponding to the effective image area on the search target image is deleted. However, in order to respond to a request to graphically display the search position on the search target image after the end of the search process, the target object search unit 40 backs up and saves the original search target image before erasure.

  The processing of S301 to S304 is performed for all pixels on the search target image. At that time, in S304, the inside of the conforming region is erased, so that matching is not performed even if the matching processing is sequentially performed on the neighboring pixels, and one target is found redundantly. Can be prevented.

  The result output in S303 is output as a search result by the search result output means 43 as described above. Specifically, the position / direction of the detected object is output as a list. In this case, for the position, the pixel position is converted into information on latitude and longitude held in the satellite image and output. It is also possible to graphically indicate the position of the target object detected on the search image. In this case, since all the searched parts are deleted, it is necessary to reload the original search target image stored in the backup. Even if the detection area erasing process of S304 is performed, if there are overlapping results, an operation is performed in which one is manually deleted and the others are deleted.

  In the above embodiment, the case where the template image is used without being updated at the time of searching has been described. However, at the time of searching, the template image is updated as in the search preparation stage. Also good. In this case, after the local matching process in S302, the second template updating unit 45 executes the same process as in S104. Specifically, the image at each angle of the template image specified by the template ID determined in S302 is updated using the pixel block at the pixel position (Xm, Ym) determined in S302. Here, as in S104, first, among the template images, the optimal angle determined in S302 and the pixel block on the search target image are synthesized. Since the mask area is not synthesized in the template image, the composition is performed only within the polygon area in the template image.

  As a specific method of synthesis, as described with reference to FIG. 7, as in S104, an integer pseudo-random number is sequentially generated for each pixel in the polygonal area, and when a predetermined number is obtained, the search target Replace with the value of the corresponding pixel in the pixel block on the image. When the update process for the optimum angle pattern is completed, the second template update unit 45 updates the other angle patterns as in S104. In the present embodiment, the second template update unit 45 functions as a part of the target object search unit 40, but its specific function is the same as the template update unit 23 in the template preparation unit 20. Actually provided as a software module. Therefore, after the processing by the local similarity determination unit 42, the template update unit 23 may execute a template image update process according to an instruction from the target object search unit 40. In this case, the template update unit 23 functions as a second template update unit.

It is a block diagram of the target object detection system which concerns on this invention. It is a figure which shows the relationship between a template image area | region and a polygonal shape. It is a figure which shows eight template images from which an angle differs. It is a flowchart which shows the process outline | summary of the calculation of a determination reference value, and the update of a template image. It is a figure which shows the mode of the process on a search object (sample) image. It is a flowchart which shows the detail of a matching process. It is a figure which shows the example which updates a template image by the synthesis | combination of a search object (sample) image and a template image. It is a figure which shows the mode of the update of a template image. It is a flowchart which shows the outline | summary of a search process. It is a figure which shows the outline of a sequential matching process. It is a flowchart which shows the detail of a sequential matching process. It is a figure which shows the optimal position of the result of the search on a search object image, and the erased area.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Search object image storage part 20 ... Template preparation part 21 ... Template preparation means 22 ... Determination reference value setting means 23 ... Template update means 30 ... Template storage part 40 ... Target object search unit 41... Sequential similarity determination means 42... Local similarity determination means 43... Search result output means 44... Detection target deletion means 45. ... Search result storage

Claims (12)

A system that detects a target from a search target image that is a search target by comparison with a template image,
Template storage means for storing the template image;
A pixel block having the same size as each template image is extracted from the search target image while sequentially moving the extraction position from the first pixel position to the last pixel position with respect to the search target image, and the template image and the pixel block are extracted. Sequential similarity judgment means for judging similarity by calculating similarity evaluation value of and comparing the similarity evaluation value with a predetermined criterion value;
When the sequential similarity determining unit determines that there is similarity, the search is performed for a pixel block having the same size as each template image for each pixel position within a predetermined range from the pixel position determined to be similar. A local similarity determination unit that extracts from a target image, calculates a similarity evaluation value between the template image and the pixel block, and determines a pixel position at which an optimal similarity evaluation value is obtained as a target object detection position;
Relates target detected by the local similarity determination means, a search result output means for outputting the pixel position location to identify the location of the pixel block in the previous SL on the search target image at that time,
Detection target erasing means for erasing the detected target by setting each pixel in the pixel block on the search target image corresponding to the detected pixel position as the target detection position to a predetermined value; , have a,
The sequential similarity determination unit and the local similarity determination unit calculate a gradation variation histogram for each of the pixels included in the template image and the corresponding pixels in the pixel block, and A target detection system , wherein a Euclidean distance is calculated as the similarity evaluation value . In claim 1,
The search result output means further outputs an ID of a template image, the target object detection system in image data. In claim 1 or claim 2 ,
In the sequential similarity determination means, a plurality of the similarity evaluation values are calculated, and when it is determined that there is no similarity with one similarity evaluation value, there is no similarity without making a determination regarding another similarity evaluation value The target object detection system characterized by being judged. In any one of Claims 1-3 ,
A polygonal effective image area is set in the template image, and the sequential similarity determination means and the local similarity determination means include a pixel group included in the effective image area in the template image, and the pixel The target detection system, wherein the similarity evaluation value is calculated based on a pixel group corresponding to the inside of the effective image area in the block. In claim 4 ,
The similarity evaluation value in the sequential similarity determination unit and the local similarity determination unit does not rotate the angle of the effective image area in the template image, and the pixel group included in the effective image area and the corresponding pixel block A similarity evaluation value calculated without depending on an angle by collating with a group of pixels within a predetermined collation method, and the effective image area while rotating the effective image area within the template image by a predetermined angle There are two types of similarity evaluation values calculated by depending on the angle by collating the pixel group included in the pixel group and the corresponding pixel group in the pixel block by a predetermined collation method. Object detection system. In claim 4 or claim 5 ,
It is defined to cut out a predetermined polygonal area specified from the sample image recording the sample of the target object to define the effective image area of the template image, and to include the effective image area even if it is rotated to an arbitrary angle. A target detection system, further comprising: a template creation unit that creates a square template image and stores the template image in the template storage unit together with information on the defined effective image area. In claim 4 of claim 6,
The search result output means corresponds to the template image ID, the pixel position specifying the position of the pixel block on the search target image, and the rotation angle of the template image for the target detected by the local similarity determination means. A target detection system, wherein an outline of an effective image area of the template image is output so as to be superimposed on the search target image. In claim 3 of claim 7,
The sequential similarity determination unit and the local similarity determination unit calculate a centroid position using the gradation values of the pixels included in the template image and the corresponding pixels in the pixel block as weights, and the centroids of both. A target detection system characterized in that a distance value between them and a difference value between both moments of inertia are used as similarity evaluation values. In claim 3 of claim 8,
The sequential similarity determination unit and the local similarity determination unit use, as the similarity evaluation value, a difference value between a gradation value of a pixel included in the template image and a corresponding pixel in the pixel block. A target detection system characterized by that. In claim 3 of claim 9,
The sequential similarity determination unit and the local similarity determination unit use, as the similarity evaluation value, a normalized correlation coefficient based on a gradation value of a pixel included in the template image and a corresponding pixel in the pixel block. A target detection system characterized by being used. In any one of Claims 1-10,
The Euclidean distance between the gradation fluctuation histograms is a Euclidean distance between a gradation fluctuation histogram based on upper bits extracted from each pixel value and a gradation fluctuation histogram based on lower bits.   The program for functioning a computer as the target object detection system in any one of Claims 1-11.

Images