JP4693742B2 - Image processing method, image processing apparatus, image processing program, and recording medium for recording image processing program - Google Patents

Description

  The present invention relates to an image processing method for performing graphic processing on image data input from an imaging apparatus or the like, an image processing apparatus, an image processing program, and a recording medium for recording the image processing program, and in particular, a work included in the image data. The present invention relates to an image processing method and an image processing apparatus for detecting a rotation angle and a position from a contour shape.

  In order to realize automation of product assembly and inspection work with high accuracy, industrial products such as food containers, printed characters, parts on printed circuit boards, and electronic parts are imaged with a camera. Therefore, an image processing apparatus that performs various image processing is required.

  In order to perform operations such as assembly and inspection using image processing, it is necessary to first detect a target work included in the image data.

  When detecting a target workpiece, one of well-known methods is analysis of a contour component indicating the shape of the target workpiece.

  In order to obtain the contour component of the target workpiece, first, a spatial differential filter is applied to the image data of the target workpiece to extract the edge component of the density change point forming the contour in the image data. In the method by the Sobel operator, which is often used as edge component extraction, an X-direction (horizontal direction) filter FX as shown in the following expression A is applied to a 3 × 3 pixel block centered on a target pixel, and Y A Sobel operator configured with a direction (vertical direction) filter FY is applied to calculate the X-direction edge strength fx and the Y-direction edge strength fy of the pixel of interest.

  The edge strength based on the calculated horizontal edge strength fx and vertical edge strength fy (the square root square sum of the strengths in two directions) is compared with a threshold value, and a pixel having a value equal to or greater than the threshold value is registered as an edge point.

  After obtaining the edge point in this way, as a method for obtaining a portion corresponding to the contour shape of the target workpiece, there is a generalized Hough transform (GHT) (see Non-Patent Document 1).

In the generalized Hough transform, first, a figure to be inspected is used as a template, and the template is considered to exist in the image after being geometrically transformed. Next, pay attention to the geometric transformation parameters that define the position and direction of the template, and vote in the parameter space. This process is roughly performed in the following steps 1 to 4.
Procedure 1 Template shape definition Procedure 2 Setting of parameter space Procedure 3 Voting Procedure 4 Peak point extraction Steps 2 to 4 of these correspond to rotated image matching.

  In the procedure 1, a template in the uv coordinate system as shown in FIG. 7 is defined, and θ, r, and α are registered in the shape definition table as shown in FIG.

In the procedure 2, first, a ballot box ([φ] [X] [Y] φ = rotation angle, X, Y = reference point, omitting parameters of size variation with few applications) is prepared. If the point Pi (i = 1, 2, 3,...) On the template is placed at the coordinates (xi, yi) in the image after the template is rotated, the image coordinates of the reference point (
X, Y) is expressed by the following equation.
X = ricos (αi + φ) + xi, Y = risin (αi + φ) + yi (1)

Here, assuming that θi, xi, and yi at the point Pi are obtained, and θ ′ at the time of shape definition table registration is θ ′, the following relationship is established.
θ = θ ′ + φ (2)

Thereafter, in the procedure 3, votes for each φ by the following procedure.
(1) Find θ ′ that satisfies Equation 2.
(2) Find r and α from θ ′ in the shape definition table.
(3) Vote for an (X, Y) cell space that satisfies Equation 1.

  Finally, the cell space (X, Y) with the largest number of votes is the peak point and the template detection position.

  Since the generalized Hough transform is said to be a process with a high calculation cost, as described in Non-Patent Document 2, there is a case in which speeding up is performed by thinning the obtained edge point in advance. is there. In addition, the image recognition method and apparatus described in Patent Document 1 obtains an approximate rotation angle and position by generalized Hough transform using a compressed image, and then restricts the rotation angle and position obtained from the compressed image. Speed-up is achieved by a two-stage process that applies generalized Hough transform.

Image Processing Textbook Editorial Committee, “Image Processing Standard Textbook”, Association for Promotion of Image Information Education, 1997, p. 254-259 Kimura Akio and Watanabe Takashi, “High-speed Generalized Hough Transform-Similar Transformation Invariant Arbitrary Shape Detection Method”, IEICE Journal D-II, IEICE, April 1998, Vol. J81-D-II, no. 4, p. 726-734 JP-A-8-123963

  Although it is possible in principle to perform rotation image matching only with the generalized Hough transform and detect the rotation angle and position, it requires a lot of calculation cost for voting processing, and the detection accuracy is also a spatial differential filter such as Sobel operator. It is hard to say that I am satisfied because I use it. Also, the voting space for translation and rotation is enormous, and some device is required.

  Therefore, in Non-Patent Document 2, the generalized Hough transform is performed on the edge points selected by the thinning process. However, in actual use, the selection by thinning may cause erroneous detection.

  In Patent Document 1, two-stage processing of generalized Hough transform is performed, but the edge direction is obtained by a spatial differential filter, and there is a problem in detection accuracy. It has been found that the success or failure of the detection accuracy in the generalized Hough transform largely depends on the result of the edge method (Satoru Igarashi and three others, “Rotation-invariant image matching based on autoregressive model of direction histogram”, precise (See Journal of Engineering Society, Japan Society for Precision Engineering, 2001, vol. 67, no. 9, p. 1519-1523), and in particular, the threshold setting has a great influence on edge detection accuracy.

  Further, when the two-stage process is used, there is a possibility that points that can be regarded as the same among the candidate points selected in the first coarse search stage may be candidates. If each subsequent process is performed on this duplicated item, time is wasted.

  An object of the present invention is to provide an image processing method, an image processing apparatus, an image processing program, and a recording medium for recording the image processing program, which improve detection accuracy and realize high-speed processing and reduction in necessary memory capacity. It is.

The present invention relates to an image processing method for detecting a target work included in image data.
An edge extraction step of extracting edge pixels from pixels constituting the image data and registering the edge direction and position;
A step of executing generalized Hough transform based on a rotation angle set for each first angle, wherein the number of votes is determined by adding the number of votes of surrounding pixels of the target pixel at the time of voting, and a predetermined designated number A coarse search step of selecting a minute pixel as a first candidate pixel;
A clustering step for performing clustering based on a distance between pixels and a rotation angle difference for the first candidate pixel, selecting a representative pixel from each cluster, and selecting a second candidate pixel;
A generalized Hough transform is performed on the second candidate pixel based on a rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate. An intermediate search step to select as pixels;
And a final search step of detecting the rotation angle and position of the target workpiece by normalized correlation with respect to the third candidate pixel.

According to the present invention, in the edge extraction step, 5 × 5 pixels are used for the thinned pixels.
It is characterized by extracting edges using Sobel operator.

In the invention, it is preferable that the first angle is 5 ° to 15 °.
In the present invention, the clustering step uses a simple clustering method, the rotation angle difference threshold is the first angle, and the inter-pixel distance threshold is the distance from the center of the reference template to the farthest edge pixel. It is a distance obtained by multiplying the distance by a predetermined ratio.

  In addition, the present invention includes a smoothing step of performing a smoothing process on the image data before the edge extraction step.

  The present invention is also an image processing program for causing a computer to execute the above image processing method.

  The present invention is also a computer-readable recording medium on which an image processing program for causing a computer to execute the above image processing method is recorded.

The present invention also provides an image processing apparatus for detecting a target work included in image data.
Edge extraction means for extracting edge pixels from pixels constituting the image data and registering edge directions and positions;
A means for executing a generalized Hough transform based on a rotation angle set for each first angle, wherein the number of votes is determined by adding the number of votes of surrounding pixels of the target pixel at the time of voting, and a predetermined designated number A coarse search means for selecting a minute pixel as a first candidate pixel;
Clustering means for performing clustering based on a distance between pixels and a rotation angle difference for the first candidate pixel, selecting a representative pixel from each cluster, and selecting a second candidate pixel;
A generalized Hough transform is performed on the second candidate pixel based on a rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate. Intermediate search means for selecting as pixels;
An image processing apparatus comprising: final search means for detecting a rotation angle and a position of a target work by normalized correlation with respect to the third candidate pixel.

  According to the present invention, in the edge extraction step, edge pixels are extracted from the pixels constituting the image data, and the edge direction and position are registered. In the coarse search step, based on the rotation angle set for each first angle. And perform generalized Hough transform. In the coarse search step, the number of votes is determined by adding the number of votes of pixels around the target pixel at the time of voting, and a predetermined number of pixels are selected as first candidate pixels.

  In the clustering step, for the first candidate pixel, clustering is performed based on a distance between pixels and a rotation angle difference, a representative pixel is selected from each cluster, and a second candidate pixel is selected. For the second candidate pixel, the generalized Hough transform is performed based on the rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate pixel. Choose as.

  In the final search step, the rotation angle and position of the target workpiece are detected by the normalized correlation for the third candidate pixel.

  As a result, the detection accuracy can be improved, the processing speed can be increased, and the required memory capacity can be reduced.

  According to the invention, in the edge extraction step, edges are extracted from the thinned pixels using a 5 × 5 pixel Sobel operator.

  Thereby, the accuracy in the edge direction is improved, and the accuracy in all subsequent steps can be improved.

According to the invention, the first angle is 5 ° to 15 °.
If the first angle is smaller than 5 °, the calculation time required for the voting process is not shortened, and if it is larger than 15 °, sufficient detection accuracy cannot be achieved.

  According to the invention, the clustering step uses a simple clustering method. At this time, the threshold value of the rotation angle difference is the first angle. The inter-pixel distance threshold is a distance obtained by multiplying the distance from the center of the reference template to the farthest edge pixel by a predetermined ratio.

  Thereby, since an appropriate cluster is generated, useless calculation is reduced and accuracy is improved.

According to the invention, a smoothing process is performed on the image data in a smoothing step before the edge extracting step.
Thereby, the extraction accuracy in the edge extraction step can be improved.

  Further, according to the present invention, it is possible to provide an image processing program for causing a computer to execute the above-described image processing method and a computer-readable recording medium on which the image processing program is recorded.

  According to the present invention, the edge extraction means extracts edge pixels from the pixels constituting the image data, registers the edge direction and position, and the coarse search means sets the rotation angle set for each first angle. Based on the generalized Hough transform. The coarse search means determines the number of votes by adding the number of votes of pixels around the target pixel at the time of voting, and selects a predetermined number of pixels as the first candidate pixels.

  The clustering means performs clustering based on the distance and rotation angle difference between the pixels for the first candidate pixel, selects a representative pixel from each cluster and selects a second candidate pixel, and an intermediate search means For the second candidate pixel, the generalized Hough transform is performed based on the rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate pixel. Choose as.

  The final search means detects the rotation angle and position of the target workpiece by the normalized correlation for the third candidate pixel.

  As a result, the detection accuracy can be improved, the processing speed can be increased, and the required memory capacity can be reduced.

In the present invention, the rotation angle and position of the target workpiece are detected by an image processing method including the following procedure.
Step 1: Smoothing processing Step 2: Edge extraction processing Step 3: Coarse search processing Step 4: Clustering processing Step 5: Intermediate search processing Step 6: Final search processing

  First, the outline of each procedure will be described based on the flowchart of image processing shown in FIG.

  In the present invention, in the rotation image collation by the generalized Hough transform, the detection accuracy is improved while the rotation angle and the position of the target workpiece are detected while increasing the speed and saving the memory.

Procedure 1: Smoothing process (Step S1)
As preprocessing for removing noise, smoothing processing using a smoothing filter is performed on image data.

Procedure 2: edge extraction process (step S2)
For the smoothed image data, the edge strength by the Sobel operator is obtained while skipping pixels. If the strength is within a preset threshold range, the edge direction and position coordinates are registered and extracted.

Procedure 3: Coarse search process (step S3)
After registering the edge direction and coordinates extracted in step 2, the generalized Hough transform is used as a rough search with a constant rotation angle increment to obtain the approximate rotation angle and position of the part corresponding to the contour shape of the target workpiece. Execute. The designated number of first candidate pixels are selected in descending order of the number of votes obtained by the generalized Hough transform.

Procedure 4: Clustering process (step S4)
Among the first candidate pixels selected in the procedure 3, there is a possibility that pixels that can be regarded as the same work are duplicated. Therefore, a plurality of pixels approximated by coordinates and rotation angles are set as the same cluster, and only a point with the largest number of votes is selected as the second candidate pixel. Other pixels are excluded from the target of subsequent processing.

Procedure 5: Intermediate search process (step S5)
For the second candidate pixel, re-voting of the generalized Hough transform is performed as an intermediate search at a rotation angle smaller than the step angle in the coarse search in the procedure 3, and the third candidate pixel is selected.

Procedure 6: final search process (step S6)
The rotation angle and coordinates obtained by the generalized Hough transform two-stage process of the procedure 3 and the procedure 5 are pixel accuracy, and an error in pixel units is expected. Therefore, the final search using the normalized correlation is performed on the third candidate pixel around the position obtained by the generalized Hough transform.

Next, each procedure will be described in detail.
Procedure 1: Smoothing process As a smoothing filter used for the smoothing process, there are an average value filter and a median filter, and an average value filter is desirable from an experimental result.

  As shown in FIG. 2, the average value filter has a pixel value of 9 pixels obtained by adding pixel values G00, G10, G20, G01, G21, G02, G12, and G22 of the surrounding 8 pixels to the pixel value G11 of the target pixel. A filter whose average is the pixel value of the target pixel.

The pixel value G (average) of the target pixel after the smoothing process is calculated by Expression 3.
G (average) = (G00 + G10 + G20 + G01 + G11 + G21 + G02 + G12 + G22) / 9 (3)

Procedure 2: Edge Extraction Processing FIG. 3 is a flowchart showing edge extraction processing.

  In step S11, the edge strength is obtained using the Sobel operator for the smoothed image data.

  The Sobel operator uses an improved operator of 5 × 5 pixels as shown in the following formula B, instead of 3 × 3 pixels.

  A Sobel operator composed of an X direction (horizontal direction) filter FX and a Y direction (vertical direction) filter FY is applied to calculate the X direction edge intensity fx and the Y direction edge intensity fy of the target pixel. If the difference value between the target pixel and the neighboring pixels adjacent to the target pixel is relatively large, the conventional 3 × 3 pixel Sobel operator represented by Expression A may be used.

The edge strength fxy is calculated by Expression 4, and the edge direction θ is calculated by Expression 5.
fxy = √ (fx ² + fy ² ) (4)
θ = tan ^-1 (fy / fx) (5)

In step S12, it is determined whether the edge strength fxy is within the range between the lower limit side threshold value TH _L and the upper limit side threshold value TH _H.

The lower limit side threshold TH _L is obtained by a discriminant analysis method for only the template region when registering the shape definition table. The discriminant analysis method is described by Hideyuki Tamura, “Computer Image Processing”, Ohmsha, July 2003, p. 140 based on the description of 140.

The upper threshold value TH _H is calculated based on statistical processing as described below.
The number of pixels having edge strength equal to or greater than the lower limit value is counted, and the number of pixels obtained by multiplying the count value by a predetermined ratio M% is calculated. The edge strength when the number of pixels having the edge strength equal to or higher than the lower limit becomes the calculated number of pixels is set as the upper limit. The ratio M% is desirably 95% to 100%.

  If the edge strength fxy is within the range, the process proceeds to step S13, and if it is out of the range, the process proceeds to step S14. In step S13, the edge direction and pixel coordinates are stored. In step S14, it is determined whether or not the processing has been completed for all the pixels. If the processing has been completed, the edge extraction processing is terminated. If not completed, the pixels are thinned out in step S15 (for example, only pixels with even addresses are processed), and the process returns to step S12.

Procedure 3: Coarse search process In procedure 3, generalized Hough transform is performed as a coarse search in increments of a predetermined rotation angle.

  FIG. 4 is a flowchart showing the rough search process. In step S21, the voting space is cleared, and voting processing is performed in step S22. In the voting process, first, the rotation angle (φ in Equation 1) is set to 0 °, ΔA, 2 × ΔA,... (360 ° −ΔA) and ΔA (unit: [°]), and Equation 2 is calculated. A satisfying θ ′ is obtained, and r and α are obtained from the position of θ ′ in the shape definition table. Then, vote for the (X, Y) cell space that satisfies Equation 1.

  In step S23, it is determined whether or not the voting process has been performed for all edges. If all have been performed, the process proceeds to step S24. If not, the process returns to step S22 for the next edge.

  In step S24, the rotation angle and position that are the maximum number of votes are searched, and the first candidate pixel is selected. The maximum number of votes is obtained for each step angle, and in step S25, it is compared with the maximum number of votes so far (the maximum number of votes that is the maximum number of votes through all angles). If it is equal to or greater than the total maximum number of votes, the process proceeds to step S26, and if it is smaller than the total maximum number of votes, the process proceeds to step S27. In step S26, the total maximum number of votes, the rotation angle, and the position are updated. For example, assuming that it is rotated by 10 °, an edge pixel of 50 ° is considered to be an edge pixel of 40 ° at the time of registration. Edge pixels that were 40 ° at the time of registration are collected from the shape definition table, and the current reference position is calculated from the relative positional relationship with the reference point and voted there.

  In step S27, it is determined whether or not processing for all rotation angles has been completed. If all have been completed, the coarse search process is terminated. If not, the process proceeds to step S28, the step angle ΔA is added to the current rotation angle, and the process returns to step S21.

  The step angle ΔA in the procedure 3 is desirably 5 ° to 15 °, and particularly desirably 8 ° to 10 °. If the step angle is smaller than 5 °, the calculation time required for the voting process is not shortened, and if it is larger than 15 °, sufficient detection accuracy cannot be achieved.

  In step S24, in order to improve the accuracy error in the edge direction obtained by the Sobel operator, the maximum number of votes is obtained by adding the number of votes of surrounding pixels. The peripheral pixels to which the number of votes is added are desirably 8 peripheral pixels in a 3 × 3 pixel centered on the target pixel.

Step 4: Clustering process From the first candidate pixels selected by the coarse search in step 3, the pixels whose position coordinates and rotation angles are approximated by clustering are collected to form the same cluster, and only the pixels with the maximum number of votes in the cluster. Is a second candidate pixel. Other pixels in one cluster are excluded from the target of subsequent processing. As clustering, simple clustering (Nearest Neighbor method-
NN method). The procedure is as follows.

  For example, as shown in the schematic diagram of FIG. 5, one data P1 is set as a cluster center, and data P2, P4, P5, etc. separated by a distance T are set as new cluster centers. P3 within the distance T belongs to the same cluster as P1.

  In this way, a cluster near the existing cluster center is a member of that cluster, and a remote cluster center is a new cluster center. Whether or not to set a new cluster center is defined in advance as a threshold value, and is set as a new center when the distance is greater than or equal to the threshold value.

  In the present invention, in order to determine whether it is a cluster member or a new cluster center, for two edge pixels, the rotation angle and the distance between the pixels are compared, and if they are separated from the threshold, Judge that it is a new cluster center.

  As the rotation angle threshold, the step angle ΔA in the coarse search is adopted, and if the rotation angles of the two edge pixels are separated from each other by the step angle or more, it is determined that they are not members of the same cluster. The distance threshold is set to a distance of S × N%, where S is the distance to the edge pixel farthest from the center among the edge pixels of the reference template. Here, the ratio N% is 70% to 100%.

  The number of candidate pixels to be clustered is set in advance, and that number of candidate pixels are selected by the coarse search. The selected candidate pixels are clustered, and the representative pixel of the upper cluster (the pixel having the maximum number of votes in one cluster) is selected as a target for subsequent processing.

Procedure 5: Intermediate Search Process In Procedure 5, generalized Hough transform is performed at predetermined rotation angle increments as an intermediate search.

  For the second candidate pixels narrowed down by clustering in procedure 4, the same search process as in the flow in procedure 3 is performed. The difference from the procedure 3 is that the candidate pixels further narrowed down by clustering are targeted, and the step angle of the rotation angle is made smaller than that in the procedure 3. If the step angle in the coarse search is 9 °, the peak angle ± 6 ° is set to 3 ° and the peak angle ± 2 ° is set to 1 °.

  After reducing the step angle, the rotation angle and coordinates of the third candidate pixels for the designated candidate number are passed to the subsequent processing in order from the pixel with the largest vote count.

Step 6: Final search process For the third candidate pixel obtained by the intermediate search in step 5, the normalized correlation is performed while rotating the captured image side around the position and rotation angle, and the rotation angle with the highest degree of coincidence is obtained. And the coordinates are output as detection results.

The normalized correlation is performed using the following correlation equation (6).
Correlation formula = {A ÷ √ (B × C)} × 100 (6)

  A = Σ (I × T) − (ΣI) × (ΣT) / N: cross-correlation between the input image and the reference image, and B = Σ (I × I) − (ΣI) × (ΣI) / N : Input image autocorrelation, C = Σ (T × T) − (ΣT) × (ΣT) / N: Reference image autocorrelation. Here, N is the number of pixels of the reference image, T is the reference image density, and I is the input image density.

  According to the present invention, in image processing for detection of a target workpiece, detection accuracy can be improved, processing speed can be increased, and required memory capacity can be reduced.

  FIG. 6 is a block diagram illustrating a configuration of the image processing system 100. The image processing system 100 includes an imaging device 101, an image processing device 102, and a display device 103, and constitutes a component position detection system as described above.

  The imaging apparatus 101 includes a CCD (charge coupled device) camera 111, an A / D (analog / digital) converter 112, a camera controller 113, a D / A converter 114, and a frame memory 115. The CCD camera 111 picks up an image of a part and outputs the amount of received light as an analog image signal. The A / D converter 112 converts the analog image signal output from the CCD camera 111 into digital data and outputs it as digital image data. The camera controller 113 stores the digital image data in the frame memory 115 for each frame and outputs it to the D / A converter 114 for display on the display device 103. The D / A converter 114 converts the digital image data output from the camera controller 113 into an analog image signal corresponding to the display device 103 and outputs the analog image signal to the display device 103. The display device 103 is realized by an LCD (Liquid Crystal Display), a CRT (Cathode Ray Tube) display, or the like, and displays an analog image signal output from the imaging device 101.

The image processing apparatus 102 includes a CPU (Central Processing Unit) 121 and a RAM (Random Access).
Memory (122), ROM (Read Only Memory) 123, and I / O (Input / Output) controller 124. The CPU 121 controls the operation of the image processing apparatus 102 based on a control program stored in the ROM 123. The image data being processed and the data being calculated are temporarily stored in the RAM 122. The I / O controller 124 is connected to an input device such as a keyboard and a mouse, a component moving device, and the like, and controls these input / output data.

  The CPU 121 and the ROM 123 constitute an edge extraction unit, a coarse search unit, a clustering unit, an intermediate search unit, and a final search unit. The CPU 121 and the ROM 123 acquire image data from the frame memory 115 via the camera controller 113 of the imaging apparatus 101. The image processing shown in the flowchart is executed. When the position and rotation angle of the target workpiece are determined by image processing, for example, it is possible to inspect whether the component is mounted in the correct direction.

  Another embodiment of the present invention is an image processing program for causing a computer to function as the image processing apparatus 102, and a computer-readable recording medium on which the image processing program is recorded. Accordingly, the image processing program and the recording medium on which the image processing program is recorded can be provided in a portable manner.

  The recording medium is read by a program reading device provided in a printer or a computer system, whereby an image processing program is executed.

As an input means of the computer system, a flat bed scanner, a film scanner, a digital camera, or the like may be used. The computer system includes these input means, a computer that executes image processing and the like by loading a predetermined program, an image display device such as a CRT display and a liquid crystal display that displays the processing results of the computer, and computer processing It consists of a printer that outputs the results to paper. Furthermore, a LAN (as a communication means for connecting to a server etc. via a network (
Local Area Network) interface etc. are provided.

  The recording medium is not limited to be read by the program reading device, and may be a microcomputer memory, for example, a ROM. The recorded program may be accessed and executed by the microprocessor, or the program read from the recording medium may be downloaded to the program storage area of the microcomputer and executed. This download function is assumed to be provided in the microcomputer in advance.

Specific examples of the recording medium include a tape system such as a magnetic tape and a cassette tape, a magnetic disk such as a flexible disk and a hard disk, and a CD-ROM (Compact Disc-
Read Only Memory) / MO (Magneto Optical) Disc / MD (Mini Disc) / DVD (
Discs of optical discs such as Digital Versatile Disc), IC (Integrated Circuit)
Cards (including memory cards) / optical cards, mask ROM, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically
Erasable Programmable Read Only Memory) and a medium that carries a fixed program including a semiconductor memory such as a flash ROM.

  In the present embodiment, the computer may have a system configuration that can be connected to a communication network including the Internet, and the image processing program may be downloaded via the communication network. In the case of downloading a program from the communication network in this way, the download function may be provided in advance in the computer or installed from another recording medium. The download program may be executed through a user interface, or may be a program that periodically downloads a program from a predetermined URL (Uniform Resource Locater).

  Compared with the case of “generalized Hough transform + normalized correlation”, which is the conventional method only, and “normalized correlation” alone, which realizes the rotated image matching processing of the present invention, the performance is improved. Confirm improvement.

(Calculation condition)
-IVS33 (manufactured by Sharp Manufacturing System Co., Ltd.) was used as the controller, standard software V3.14 was used for normalization correlation, and other software was used for the others.
・ Search area size 512 × 480,
・ Rectangle template 96 × 96
The processing by “generalized Hough transform + normalized correlation” is an example, and the processing by only “generalized Hough transform” and only by “normalized correlation (rotation angle 10 ° increments)” are referred to as comparative examples 1 and 2, respectively. Further, as a reference example, processing based only on “normalized correlation (rotation angle in increments of 20 °)” was also performed.

In Comparative Example 2 and Reference Example, 36 templates of 10 ° units were prepared.
First, the processing speed results are shown in Table 1.

The time required for processing in Example 1 was set to 1.00, and the processing speed was evaluated based on the ratio.
Next, the detection accuracy results are shown in Table 2.

  As for the processing speed, the result of Comparative Example 1 was better than that of the Example, but the Example greatly exceeded the detection accuracy. Therefore, if it judged comprehensively, it became clear that the Example was superior to Comparative Examples 1 and 2.

It is a flowchart which shows the image processing of this invention. It is a figure which shows the example of an average value filter. It is a flowchart which shows an edge extraction process. It is a flowchart which shows a rough search process. It is a schematic diagram for demonstrating a clustering process. 1 is a block diagram showing a configuration of an image processing system 100. FIG. It is a figure which shows the example of a definition of a shape template. It is a figure which shows the example of a shape definition table | surface.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Image processing system 101 Imaging apparatus 102 Image processing apparatus 103 Display apparatus 111 CCD camera 112 A / D converter 113 Camera controller 114 D / A converter 115 Frame memory 121 CPU
122 RAM
123 ROM
124 I / O controller

Claims (8)

In an image processing method for detecting a target work included in image data,
An edge extraction step of extracting edge pixels from pixels constituting the image data and registering the edge direction and position;
A step of executing generalized Hough transform based on a rotation angle set for each first angle, wherein the number of votes is determined by adding the number of votes of surrounding pixels of the target pixel at the time of voting, and a predetermined designated number A coarse search step of selecting a minute pixel as a first candidate pixel;
A clustering step for performing clustering based on a distance between pixels and a rotation angle difference for the first candidate pixel, selecting a representative pixel from each cluster, and selecting a second candidate pixel;
A generalized Hough transform is performed on the second candidate pixel based on a rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate. An intermediate search step to select as pixels;
And a final search step of detecting the rotation angle and position of the target workpiece by normalized correlation with respect to the third candidate pixel.   2. The image processing method according to claim 1, wherein in the edge extraction step, edges are extracted from the thinned pixels using a 5 × 5 pixel Sobel operator.   The image processing method according to claim 1, wherein the first angle is 5 ° to 15 °.   The clustering step uses a simple clustering method, the rotation angle difference threshold is the first angle, and the inter-pixel distance threshold is a predetermined ratio to the distance from the center of the reference template to the farthest edge pixel. The image processing method according to claim 1, wherein the distance is multiplied by.   The image processing method according to claim 1, further comprising a smoothing step of performing a smoothing process on the image data before the edge extraction step.   An image processing program for causing a computer to execute the image processing method according to claim 1.   A computer-readable recording medium storing an image processing program for causing a computer to execute the image processing method according to claim 1. In an image processing apparatus for detecting a target work included in image data,
Edge extraction means for extracting edge pixels from pixels constituting the image data and registering edge directions and positions;
A means for executing a generalized Hough transform based on a rotation angle set for each first angle, wherein the number of votes is determined by adding the number of votes of surrounding pixels of the target pixel at the time of voting, and a predetermined designated number A coarse search means for selecting a minute pixel as a first candidate pixel;
Clustering means for performing clustering based on a distance between pixels and a rotation angle difference for the first candidate pixel, selecting a representative pixel from each cluster, and selecting a second candidate pixel;
A generalized Hough transform is performed on the second candidate pixel based on a rotation angle set for each second angle smaller than the first angle, and a predetermined number of pixels are determined as the third candidate. Intermediate search means for selecting as pixels;
An image processing apparatus comprising: a final search unit that detects a rotation angle and a position of a target workpiece by using a normalized correlation for the third candidate pixel.

Images