JP2011058925A - Method and device for inspecting image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image inspection method for precisely inspecting an image to be inspected even in a case that the steep change of feature quantity is produced in the image to be inspected, and an image inspection device. <P>SOLUTION: A good image is preliminarily divided into a plurality of good divided images to calculate MAHALANOBIS distance at each good divided image to form a reference space to thereby set a quality determining reference value for determining the quality of the image to be inspected. The image to be inspected is obtained (S41) to apply image processing (S42). Subsequently, the image to be inspected is divided into a plurality of the divided images to be inspected (S43). Next, the predetermined feature quantities of the individual divided images to be inspected are extracted (S44). Then, the MAHALANOBIS distances of the respective divided images to be inspected are calculated to form a signal space (S45). The MAHALANOBIS distances of the corresponding to the signal space is compared with the quality determining reference value in the good divided images present at the same position at each divided image to be inspected to determine the quality of the image to be inspected (S46-S48). <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an image inspection method and an image inspection apparatus for determining whether or not an image to be inspected that shows the appearance of a product in order to determine whether there is a defect in the product to be inspected, for example.

  2. Description of the Related Art Conventionally, a method is known in which the presence or absence of defects in a product to be inspected is determined from an image to be inspected by photographing the appearance of the product. As one of the methods, there is a method for inspecting the quality of an inspected image using the Mahalanobis Taguchi method. The basic procedure for image inspection by the Mahalanobis Taguchi method is as follows.

1) Feature values are extracted from a plurality of product images in the most normal state according to a plurality of items representing the state. In this way, a data group of normal products and non-defective products is referred to as a reference space, and becomes a reference when performing determination, diagnosis, and prediction by the Mahalanobis Taguchi method.
2) Each feature value is normalized using the average value and standard deviation of the feature values calculated for each item.
3) Calculate all correlations between the standardized feature quantities and create a correlation matrix.
4) The Mahalanobis distance of the reference space, which is a group of products in the most normal state, is calculated using the normalized feature quantity and the inverse matrix of the correlation matrix. The average value of the Mahalanobis distance in this reference space is about 1.
5) The Mahalanobis distance is calculated using the same procedure as above for the product image (inspected image) whose state is unknown.
6) Based on a preset threshold value of hamaranobis distance (good / bad determination reference value), the state of the product whose state is unknown is determined. In the Mahalanobis Taguchi method, the Mahalanobis distance increases as the degree of abnormality increases. Therefore, if the Mahalanobis distance becomes larger than the reference value, it is determined that the product is in an abnormal state (defective). In this way, data that does not belong to the reference space is called a signal space, and its state is determined, diagnosed, and predicted based on the Mahalanobis distance.

  In this way, the Mahalanobis-Taguchi method uses a multivariate analysis method to determine the state of a product, so that various information (features) can be obtained in the same manner as a visual inspection in which an operator visually determines the state of a product. Comprehensive judgment based on this can be made. Various methods have been proposed in order to accurately inspect the quality (for example, see Patent Documents 1 and 2).

  For example, Patent Document 1 discloses that primary parameters such as luminance and density, average values of primary parameters, and standard deviations are used as image feature amounts used for calculation of Mahalanobis distance in order to accurately inspect the quality of the entire image to be inspected. A feature quantity distribution such as a luminance distribution of the entire image to be inspected is used instead of a secondary parameter such as. This is because the Mahalanobis distance is calculated based on the feature amount including the information on the entire image to be inspected, so that the quality of the entire image to be inspected can be determined more accurately than when inspecting with some feature amounts of the image to be inspected. The aim is to be able to inspect.

  Patent Document 2 discloses that an image to be inspected is divided into small regions, feature amounts are extracted from each small region, and a statistical distance measure (Mahalanobis distance, etc.) between the extracted feature amount and a predetermined feature amount reference value. Describes a method for inspecting the quality of an image to be inspected by using a plurality of sizes of small areas to be divided. This aims at accurately detecting not only local defects but also global defects.

JP-A-2005-252451 JP 2007-132757 A

  However, since the method described in Patent Document 1 calculates the Mahalanobis distance based on the luminance distribution, this method is used for the appearance inspection of products, particularly defects in products with a large change in appearance and steep feature amounts in the appearance. When used for the appearance inspection of defects in products with various changes, there is a problem that the feature quantity of the defects is hidden in the noise and cannot be inspected with high accuracy. In addition, since the position of the defect cannot be specified, there is a problem that the inspection result cannot be effectively used secondarily, for example, the inspection result is fed back to the previous process.

  In addition, although the method described in Patent Document 2 has a defect, the Mahalanobis distance may not be determined because the Mahalanobis distance is small. Therefore, a second determination is added in addition to the determination based on the Mahalanobis distance. is doing. For this reason, there is a problem that the inspection method becomes complicated.

  The present invention has been made in view of the above problems, and can be easily inspected without adding the second determination, and can be inspected accurately even when there is a sharp change in the feature amount in the inspected image. It is a first object to provide an image inspection method and an image inspection apparatus that can be used.

  It is a second object of the present invention to provide an image inspection method and an image inspection apparatus that can feed back an inspection result to a previous process.

In order to solve the above problems, the present invention is an image inspection method for determining whether an image is good or bad,
A reference space acquisition step of acquiring a reference space composed of Mahalanobis distances for each of the good product divided images when the good product image is divided into a plurality of images (hereinafter referred to as “good product split images”);
Inspected image acquisition step of acquiring the inspected image;
An inspected image dividing step for dividing the inspected image into a plurality of images (hereinafter referred to as “inspected divided images”);
A signal space creating step for creating a signal space consisting of Mahalanobis distance for each of the inspected divided images based on the feature amount of the inspected divided image and the feature amount of the non-defective divided image;
By comparing the Mahalanobis distance of the corresponding signal space for each of the inspected divided images with a pass / fail criterion value determined based on the Mahalanobis distance of the reference space in the good product divided image at the same position. A pass / fail determination step for determining pass / fail of the image.

  According to this, since the inspected image is divided into the inspected divided images, when there is a defect in the inspected image, the ratio of the defects in the inspected divided image is occupied by the defects in the entire inspected image. Will be greater than the percentage. And since the pass / fail determination is performed for each divided image to be inspected, it is possible to easily detect the defect even when inspecting the defect of the product having a sharp change in the feature amount in the appearance. Therefore, the inspection can be performed with high accuracy.

  In addition, since the signal space consisting of the Mahalanobis distance for each divided image to be inspected is created based on the feature amount of the inspected image and the feature amount of the non-defective divided image, the Mahalanobis distance of the signal space It can be said that the difference in the feature amount between the image and the non-defective product divided image is reflected. Therefore, it is reduced that the defect is not determined because the Mahalanobis distance is small even though there is a defect. Therefore, since it is not necessary to add the second determination in addition to the determination based on the Mahalanobis distance, the inspection can be easily performed.

  Further, the image inspection method of the present invention includes a defect position recording step of recording a position on the inspection image of the inspection divided image determined to be defective in the quality determination step.

  Accordingly, the inspection result can be fed back to the previous process by referring to the recorded defect position.

In the image inspection method of the present invention, the reference space acquisition step includes
A non-defective image acquiring step for acquiring the non-defective image;
A non-defective image dividing step for dividing the non-defective image into the non-defective divided images;
And a reference space creating step of creating the reference space by calculating a Mahalanobis distance for each of the non-defective product divided images.

  As a result, a reference space corresponding to the signal space can be acquired.

In the image inspection method of the present invention, the non-defective image acquisition step acquires a plurality of non-defective images,
The reference space creation step includes
A first extraction step of extracting a plurality of predetermined feature amounts for each non-defective product divided image;
A first normalization step for normalizing those feature values and calculating good product standard feature values;
A first matrix calculating step of calculating a non-defective product correlation coefficient matrix comprising correlation coefficients between the non-defective product standard feature values for each of the non-defective product divided images at the same position between the plurality of non-defective products images;
A first distance calculating step of calculating a Mahalanobis distance of the reference space on the basis of the good product criterion feature quantity and the good product correlation coefficient matrix for each good product divided image.

  As a result, the reference space can be created and, at that time, the reference space is created using a plurality of non-defective images, so that the reliability of the reference space can be increased.

  Further, in the image inspection method of the present invention, based on the maximum value among the Mahalanobis distances of the good product divided images at the same position among the plurality of good product images, the pass / fail judgment reference value for each good product divided image is determined. A reference value setting step for setting is included.

  As described above, since the pass / fail judgment reference value is set based on the maximum value among the Mahalanobis distances of the good product split images located at the same position among a plurality of good product images, the accuracy of the pass / fail judgment can be further improved. .

In the image inspection method of the present invention, the signal space creating step includes
A second extraction step for extracting the same feature quantity as the feature quantity extracted in the first extraction step for each of the inspected divided images;
A second normalization step of normalizing the feature amount extracted in the second extraction step with reference to the feature amount of the non-defective product divided image to calculate an inspection standard feature amount;
A second distance calculating step of calculating a Mahalanobis distance of the signal space based on the inspected reference feature quantity and the non-defective product correlation coefficient matrix.

  In this way, the inspection standard feature quantity is calculated based on the feature quantity of the non-defective product divided image, and the Mahalanobis distance of the signal space is calculated based on the inspection standard feature quantity and the non-defective product correlation coefficient matrix. Thus, it is possible to create a signal space in which the difference between the feature amounts of the inspected divided image and the non-defective divided image is reflected.

  In the image inspection method of the present invention, when the first and second extraction steps draw a specific horizontal line with respect to the luminance waveform of each image from which the feature amount is acquired, the horizontal line and the luminance waveform are drawn. And at least one of a differential value that is the number of intersections with each other, an integral value that is an interval of the luminance waveform above the horizontal line, and a difference between the maximum luminance value and the minimum luminance value as the feature amount.

  According to this, the differential value, which is the number of intersections between the horizontal line and the luminance waveform, can be used as an index of the number of changes in the image feature amount. In other words, it can be said that the larger the differential value, the more portions in the image where the feature amount varies. Therefore, by using this differential value as a feature value, even in an image that includes many parts where the feature value varies, it is possible to create a reference space or a signal space that accurately reflects the feature value. The quality of the image to be inspected can be determined with high accuracy.

  Further, the integral value, which is the section of the luminance waveform above the horizontal line, can be used as an index of the ratio of the portion having a certain feature amount or more in the image. In other words, it can be said that the larger the integral value, the more parts in the image that have a certain or larger feature amount. Therefore, by using this integral value as a feature amount, even in an image including many parts having a feature amount above a certain level, a reference space or signal space in which the feature amount is accurately reflected is created. Therefore, the quality of the image to be inspected can be determined with high accuracy.

  Further, the difference between the maximum luminance value and the minimum luminance value can be used as an indicator of the magnitude of the variation of the feature amount in the image. In other words, it can be said that as the difference between the maximum luminance value and the minimum luminance value is larger, a portion where the feature amount fluctuates greatly is included in the image. Therefore, by using this difference as a feature value, a reference space or signal space in which the feature value is accurately reflected can be created even for an image including a part where the feature value greatly varies. The quality of the inspection image can be determined with high accuracy.

  Further, the position of the horizontal line may be variable. As a result, even in an image in which the feature amount is complicatedly changed, it is possible to create a reference space or a signal space in which the feature amount is accurately reflected. Therefore, it is possible to accurately determine the quality of the image to be inspected.

Further, in the image inspection method of the present invention, an inspected image processing step of performing predetermined image processing on the inspected image in order to easily detect defects in the inspected image;
A non-defective image processing step for performing the same image processing as the image processing on the non-defective image.

  According to this, since the image to be inspected is subjected to predetermined image processing, when there is a defect in the image to be inspected, it is possible to better extract the feature amount of the defect. Further, the same image processing is performed on the non-defective image, and the reference space and the signal space can be created under the same conditions. Therefore, the quality of the inspected image can be determined with high accuracy.

  In the image inspection method of the present invention, the non-defective image acquisition step is a step of acquiring an image included in an image group having the highest appearance frequency as the non-defective image.

  According to this, since the image group having the highest appearance frequency is a non-defective image, it can be said that the reliability as the non-defective image is high.

In the image inspection method of the present invention, the image to be inspected is an image showing an appearance of a product to be visually inspected,
The appearance of the product is an appearance in which the feature amount fluctuates greatly to the extent that the feature amount of the assumed defect is hidden.

  As described above, it is preferable to apply the present invention to the product inspection of the appearance whose feature quantity greatly varies.

The present invention is an image inspection apparatus for determining the quality of an image,
Reference space acquisition means for acquiring a reference space consisting of the Mahalanobis distance for each good product divided image when the good product image is divided into a plurality of images (hereinafter referred to as “good product divided images”);
Inspected image acquisition means for acquiring the inspected image;
Inspected image dividing means for dividing the inspected image into a plurality of images (hereinafter referred to as “inspected divided images”);
Based on the feature amount of the inspected divided image and the feature amount of the non-defective divided image, signal space creating means for creating a signal space consisting of the Mahalanobis distance for each of the inspected divided images;
By comparing the Mahalanobis distance of the corresponding signal space for each of the inspected divided images with a pass / fail criterion value determined based on the Mahalanobis distance of the reference space in the good product divided image at the same position. And a pass / fail judgment means for judging pass / fail of the image.

  According to this, the same effect as the image inspection method of the present invention can be obtained.

1 is an external view of an image inspection apparatus 1. FIG. FIG. 4 is a view showing a surface of a work 4. It is a basic flowchart for creating a reference space. It is a basic flowchart for calculating a Mahalanobis distance of an image to be inspected and performing pass / fail judgment. It is a detailed flowchart for creating a reference space. It is a figure for demonstrating the division | segmentation method of the good quality image 41. FIG. It is the figure which showed the good quality division | segmentation image 411 in the same position between the some good quality images 41. FIG. It is a detailed flowchart for calculating the Mahalanobis distance of the image to be inspected and performing pass / fail judgment. As an example of the image to be inspected, it is a diagram showing an inspected image 42 that is a defective product and an inspected image 43 that is a non-defective product.

  The best mode for carrying out the present invention will be described below with reference to the drawings. In FIG. 1, the external view of the image inspection apparatus 1 of this invention is shown. The image inspection apparatus 1 is an apparatus that determines the quality of the workpiece 4 by determining whether the workpiece 4 shown in FIG. As shown in FIG. 1, the image inspection apparatus 1 includes a camera 2, an illumination 3, an electronic calculator 5, and a display device 6.

  The camera 2 images the external appearance of the work 4, and a CCD camera, a CMOS camera, an analog camera, or the like can be used. In this example, a CCD camera is used as the camera 2.

  The illumination 3 is used to adjust the brightness of the imaged portion of the workpiece 4 when the camera 2 captures the appearance of the workpiece 4, and fiber illumination, LED illumination, ring illumination, coaxial epi-illumination, or the like can be used. it can. In this example, LED ring illumination is used as the illumination 3. The image data picked up by the camera 2 is sent to the electronic computer 5, and after the data processing is performed by the electronic computer 5, pass / fail judgment is performed, and the result is displayed on the display device 6.

  On the other hand, the workpiece 4 is a product to be inspected. In this example, the workpiece 4 is a cut metal part, and is used to inspect the surface of the workpiece 4 for defects such as scratches, burrs, and foreign matters. FIG. 2 is a view showing the surface of the workpiece 4. As shown in FIG. 2, the inspection target workpiece 4 exemplified here includes portions having various characteristics 450 (for example, patterns and shapes) in the appearance. Therefore, the work has a sharp change in the feature amount indicated by the feature 450 in the appearance.

  Then, the image inspection apparatus 1 determines the quality of the workpiece 4 using the Mahalanobis Taguchi method. Here, FIGS. 3 and 4 show flowcharts of basic data processing performed by the electronic computer 5. FIG. 3 is a flowchart for calculating the Mahalanobis distance of a plurality of product images in the most normal state, that is, a non-defective image group having the highest appearance frequency, that is, for creating a reference space. First, a good original image is input, and image processing such as filtering and edge enhancement is performed on the input image as necessary (S11). Then, after the input image is divided (S12), the feature amount is extracted from each divided image (S13), and the Mahalanobis distance is calculated for each divided image at the same position among the plurality of images (S14).

  On the other hand, FIG. 4 is a flowchart for calculating the Mahalanobis distance of the image to be inspected and making a pass / fail judgment. First, an inspected image of an inspection product is input, and if necessary, image processing such as filtering and edge enhancement is performed on the input image to facilitate detection of defects (S21). Then, the image to be inspected is divided (S22), the feature amount is extracted from each divided image (S23), and the Mahalanobis distance is calculated (S24). The calculated Mahalanobis distance is compared with a reference value for pass / fail determination (S25). If the calculated Mahalanobis distance is smaller than the reference value, it is determined to be a non-defective product, and the result is displayed on the display device 6 (S26). Thereafter, the process of the flowchart of FIG. On the other hand, if the calculated Mahalanobis distance is larger than the reference value, it is determined as a defective product (S27), the position where the defect exists is specified, and the information is stored in a storage device (not shown) of the electronic computer 5. The result is displayed on the display device 6 (S28). Thereafter, the process of the flowchart of FIG.

  As described above, the present invention determines pass / fail for each divided image. And this invention has the characteristics in the calculation method of Mahalanobis distance. Hereinafter, the details of the Mahalanobis distance calculation method of the present invention will be described.

  First, a method for creating a reference space will be described with reference to the flowchart shown in FIG. 5 and FIG. First, product images of a plurality of workpieces 4 in the most normal state are captured by the camera 2, and the image data is stored in a storage device (not shown) of the electronic computer 5 (S31). Here, the product image in the most normal state is an image of a good product group having the highest appearance frequency in the case of image inspection. This is because if the appearance frequency is high, it can be said that the probability that the work 4 is normal and non-defective is high. In addition, the number of non-defective images to be inputted is inputted and n pieces of non-defective images are inputted. Step S31 corresponds to the “non-defective image acquisition step” of the present invention.

  Then, image processing such as filtering and edge enhancement is performed on each non-defective image input in S31 (S32). Here, edge enhancement processing is performed. Originally, this image processing is intended to make it easier to detect defects when there is a defect in the image to be inspected, but it is necessary to perform inspection under the same conditions regardless of whether the product is non-defective or defective. Image processing is also applied to non-defective images that are known not to exist. Step S32 corresponds to the “non-defective image processing step” of the present invention.

  Next, as shown in FIG. 6, the individual non-defective images 41 subjected to the image processing are divided into arbitrary sizes and arbitrary sizes (S33). The division direction may be determined in consideration of the defect generation direction and the division size in consideration of the size of the defect to be detected. Here, however, the original image 41 of 210 × 1320 pixels is converted to the non-defective product divided image 411 of 10 × 1320 pixels ( 1) to 411 (N). That is, the non-defective image 41 is divided into N parts in the vertical width. Step S33 corresponds to the “non-defective image division step” of the present invention.

  Next, each of the non-defective product divided images 411 (1) to 411 (N) is subjected to arithmetic processing, and a feature amount is extracted based on a predetermined item (S34). This item needs to represent the state of the work 4, but here, when an arbitrary horizontal line is drawn with respect to the luminance waveforms of the respective non-defective product divided images 411 (1) to 411 (N), the horizontal line and the luminance waveform are displayed. The feature amount based on the luminance information is used, such as a differential value that is the number of intersections of, an integral value that is the sum of the sections of the luminance waveform that are above the horizontal line, and a difference between the maximum luminance value and the minimum luminance value.

  Here, the differential value, which is the number of intersections between the horizontal line and the luminance waveform, can be used as an index of the number of changes in the image feature amount. In other words, it can be said that the larger the differential value, the more portions in the image where the feature amount varies. Therefore, by using this differential value as a feature value, even in an image that includes many parts where the feature value varies, it is possible to create a reference space or a signal space that accurately reflects the feature value. It is considered that the quality of the inspected image can be determined with high accuracy.

  Further, the integral value, which is the section of the luminance waveform above the horizontal line, can be used as an index of the ratio of the portion having a certain feature amount or more in the image. In other words, it can be said that the larger the integral value, the more parts in the image that have a certain or larger feature amount. Therefore, by using this integral value as a feature amount, even in an image including many parts having a feature amount above a certain level, a reference space or signal space in which the feature amount is accurately reflected is created. Therefore, it is considered that the quality of the inspected image can be determined with high accuracy.

  Further, the difference between the maximum luminance value and the minimum luminance value can be used as an indicator of the magnitude of the variation of the feature amount in the image. In other words, it can be said that as the difference between the maximum luminance value and the minimum luminance value is larger, a portion where the feature amount fluctuates greatly is included in the image. Therefore, by using this difference as a feature value, a reference space or signal space in which the feature value is accurately reflected can be created even for an image including a part where the feature value greatly varies. It is considered that the quality of the inspection image can be determined with high accuracy.

  In S34, the feature amount is extracted with the position of the horizontal line being variable. As a result, even in an image in which the feature amount fluctuates in a complicated manner, the feature amount can be accurately extracted.

Then, as a result of extracting the feature quantity X _gij from the respective non-defective product divided images 411 (1) to 411 (N) in S34, the result is as shown in Table 1.

Table 1 shows the feature values X _gij of the respective non-defective product divided images 411 (1) to 411 (N) for each of the n non-defective products 41. The extracted feature amount X _gij has k types (the luminance waveform differential value, the luminance waveform integral value, the difference between the maximum luminance value and the minimum luminance value, etc.), and Table 1 shows the characteristic amount for each item. X _gij is shown. For example, as shown in Table 1, the feature quantity of the non-defective image 41 having the image No. “b”, the divided image No. “1”, and the item No. “2” is “X _b12 ”. That is, in Table 1, among the subscripts attached to the feature amount X _gij , the leftmost character g indicates the image No. of the non-defective image 41, and the middle character i indicates the No of the non-defective product divided image 411, The rightmost character j indicates item No. Hereinafter, the character g is used as the image No. of the non-defective image 41, the character i is used as the No of the non-defective product divided image 411, and the character j is used as the item No. Step S34 corresponds to the “first extraction step” of the present invention.

Next, the Mahalanobis distance of each of the non-defective product divided images 411 (1) to 411 (N) is calculated to create a basic space (S35). Specifically, the average value x _gj and the standard deviation σ _gj of the feature amount X _gij are calculated for each item j for each good product image 41. For example, the average value x _gj and the standard deviation σ _gj of the feature amount X _bi1 of the non-defective image 41 with the image No. g = “b” and the item No. j = “1” are the image No. “b”, Table 1 _Calculation is performed using the feature amounts X _{b11 to} X _bn1 of the item No “1”.

Then, the calculated average value x _gj and standard deviation σ _gj are substituted into the following equation (1) to normalize the feature amount X _gij (S35).

The standardized feature quantity x _gij (good product standard feature quantity) is as shown in Table 2. In Table 2, the feature quantities x _gij normalized for each of the divided images 411 (1) to 411 (N) at the same position among the plurality of non-defective images 41 are arranged.

For example, the divided image No of each good product image 41 is “1”, and the feature quantity x _gij of the item “1” is x _a11 , x _b11 ,..., X _n11 as shown in Table 2.

Then, as shown in FIG. 7, for each of the non-defective product divided images 411 (1) to 411 (N) in the same position among the non-defective product images 41, the relationship between the item p and the item q is expressed by the following equation (2). The number r _pq is calculated (S35). Then, the correlation coefficient r _pq is matrixed as in the following equation (3) to calculate a correlation coefficient matrix R (S35).

For example, the correlation coefficient _{r 12} scores 1 in good divided image 411 (1) (p = 1) and Item 2 (q = 2), the above equation (2), to be calculated by the following equation (4) Become. That is, in this case, the correlation coefficient r ₁₂ is calculated using the feature amounts x _gij of the items No. “1” and “2” in the divided image No. “1” in Table 2.

Thus, non-defective divided image 411 (1) ~411 (N) each at by calculating the correlation coefficient _{r pq,} good divided image 411 (1) ~411 (N) per the formula (3) shows a phase relationship A number matrix R is calculated.

Then, the Mahalanobis distance MD _ig of the g-th non-defective product image 41 of the i-th non-defective product divided image 411 (i) is calculated by the following equation (5). In Equation (5), A is an inverse matrix of the correlation coefficient matrix R of the non-defective product divided image 411 (i).

Then, the Mahalanobis distance MD _ig is calculated with respect to all the non-defective product divided images 411 of all the non-defective products images 41 using Equation 5 (S35). As a result of calculating the Mahalanobis distance MD _ig for all the non-defective product divided images 411, the results are as shown in Table 3 here. In Table 3, the maximum value of the Mahalanobis distance _{MD ig} of each non-defective divided image 411 is also shown, for example, the maximum value of the Mahalanobis distance _{MD ig} non-defective divided image 411 (1) is "1.1" ing.

  Step S35 corresponds to the “first normalization step”, “first matrix calculation step”, and “first distance calculation step” of the present invention. Steps S34 and S35 correspond to the “reference space creation step” of the present invention. Steps S31 to S35 correspond to the “reference space acquisition step” of the present invention, and the electronic computer 5 that executes steps S31 to S35 corresponds to “reference space acquisition means” of the present invention.

Next, based on the Mahalanobis distance MD _ig calculated in S35, a Mahalanobis distance reference value (non-defective product determination reference value) that serves as a reference for determining pass / fail of the inspected image is set for each non-defective product divided image 411 (S36). Specifically, the Mahalanobis distance reference value was set as shown in Table 4 with reference to the maximum value of the Mahalanobis distance MD _ig in each non-defective product divided image 411. In Table 4, the Mahalanobis distance reference value = the maximum value of the Mahalanobis distance _MDig + 0.3. Thereafter, this Mahalanobis distance reference value is stored in a storage device (not shown) of the electronic computer 5, and the processing of the flowchart of FIG. Step S36 corresponds to the “reference value setting step” of the present invention.

  Next, a method for determining pass / fail by calculating the Mahalanobis distance of the image to be inspected will be described. Here, FIG. 8 is a flowchart showing the processing executed by the electronic computer 5 when determining the quality of the inspected image. The method for calculating the Mahalanobis distance of the image to be inspected is that there is one input image, the average value for each item used for normalization, the standard deviation, and the correlation coefficient matrix R used for calculating the Mahalanobis distance. The inverse matrix A is the same as the method for creating the reference space except that the value at the time of creating the reference space is used. Based on this point, a method for determining pass / fail of an inspected image will be described below with reference to the flowchart of FIG.

  First, a product image of the workpiece 4 to be inspected is picked up by the camera 2 as an image to be inspected, and the image data is stored in a storage device (not shown) of the electronic computer 5 (S41). Step S41 corresponds to the “inspected image acquisition step” of the present invention, and the electronic computer 5 that executes step S41 corresponds to the “inspected image acquisition means” of the present invention.

  Then, image processing such as filtering and edge enhancement is performed on the inspection image input in S41 (S42). Step S42 corresponds to the “inspected image processing step” of the present invention.

  Next, the inspected image subjected to the image processing is divided in the same direction as the non-defective image 41 with the same size (S43). That is, the 210 × 1320 pixel inspection image is divided into 10 × 1320 pixel inspection divided images. Step S43 corresponds to the “inspected image dividing step” of the present invention, and the electronic computer 5 that executes step S43 corresponds to the “inspected image dividing means” of the present invention.

Next, an arithmetic process is performed on each divided image to be inspected, and a feature amount Y _ij is extracted based on the same items as when the reference space is created (S44). That is, when an arbitrary horizontal line is drawn on the luminance waveform of the divided image to be inspected, the differential value that is the number of intersections of the horizontal line and the luminance waveform, the integral value that is the sum of the intervals of the luminance waveform that is above the horizontal line, and the maximum luminance A feature amount Y _ij based on luminance information such as a difference between the value and the minimum luminance value is extracted. Step S44 corresponds to the “second extraction step” of the present invention.

Next, the Mahalanobis distance of each divided image to be inspected is calculated by the same method (S35) as that for creating the reference space, and a signal space is created (S45). Specifically, the feature amount Y _ij for each item j for each divided image to be inspected is normalized (inspected reference feature amount) by the following equation (6) (S45). At this time, the average value x _j and the standard deviation σ _j when the reference space is created are used.

Then, the Mahalanobis distance MD ′ _i of the i-th inspection divided image is calculated by the following equation (7). At this time, the inverse matrix A of the correlation coefficient matrix R is the one used when the reference space is created.

  Step S45 corresponds to the “second normalization step” and the “second distance calculation step” of the present invention. Steps S44 and S45 correspond to the “signal space creation step” of the present invention, and the electronic computer 5 that executes steps S44 and S45 corresponds to the “signal space creation means” of the present invention.

Next, each calculated Mahalanobis distance MD ′ _i is compared with the corresponding reference value for pass / fail judgment in Table 4 (S46). Then, all of the Mahalanobis distance MD _'i is judged to be good is smaller than the reference value, the display result on the display device 6 (S47). Then, the process of the flowchart of FIG. 8 is complete | finished. On the other hand, if at least one of the calculated Mahalanobis distances MD ′ _i is larger than the reference value, it is determined as a defective product (S48), the position where the defect exists is specified, and the information is stored in the storage device (invalid) of the electronic computer 5. In addition, the result is displayed on the display device 6 (S49). As information for specifying the position where the defect exists, for example, No of the divided image to be inspected corresponding to the Mahalanobis distance MD ′ _i determined to be larger than the reference value is used. Then, the process of the flowchart of FIG. 8 is complete | finished. Steps S46 to S48 correspond to the “good / bad determination step” of the present invention, and the computer 5 that executes steps S46 to S48 corresponds to the “good / bad determination unit” of the present invention. Step S49 corresponds to the “defect position recording step” of the present invention.

  Here, the quality determination result when the images to be inspected are inspected images 42 and 43 as shown in FIG. 9 will be described. The inspection image 42 shown in FIG. 9A has a defect 70 and is a defective product. Then, as shown in FIG. 9A, the defect 70 is included in the inspected divided images 421 (1) to (N) of the inspected image 42, and in particular, the inspected divided image 421 (10). It is assumed that many are included. On the other hand, the inspected image 43 shown in FIG. Therefore, the inspection divided images 431 (1) to (N) of the inspection image 43 naturally do not include defects.

  When the Mahalanobis distance was calculated for each of the inspected divided images 421 and 431 with respect to the inspected images 42 and 43, it was as shown in Table 5.

  As shown in Table 5, in the inspected image 42, the divided image No. 10, there is a defect 70 at a position in the area 10. The Mahalanobis distance of 10 is divided image No. It is larger than the reference value 1.9 of 10. Therefore, it was determined that the inspected image 42 is a defective product. On the other hand, the inspected image 43 was determined to be a non-defective product because the Mahalanobis distance of all the inspected divided images 431 was less than or equal to the reference value shown in Table 4.

  As described above, in the present invention, the image to be inspected is divided into a plurality of divided images to be inspected. Therefore, when there is a defect in the image to be inspected, the ratio of the defects in the divided image to be inspected is determined as follows. This is larger than the ratio of defects in the entire inspection image. And since the pass / fail determination is performed for each divided image to be inspected, it is possible to easily detect the defect even when inspecting the defect of the product having a sharp change in the feature amount in the appearance. Therefore, the inspection can be performed with high accuracy.

Further, since the signal space consisting of the Mahalanobis distance for each divided image to be inspected is created using the average value x _j , standard deviation σ _j , and inverse matrix A when the reference space is created, the Mahalanobis of the signal space is created. It can be said that the distance MD ′ _i reflects the difference between the feature amounts of the inspected divided image and the non-defective divided image. Therefore, it is reduced that the defect is not determined because the Mahalanobis distance is small even though there is a defect. Therefore, since it is not necessary to add the second determination in addition to the determination based on the Mahalanobis distance, the inspection can be easily performed.

  Further, in the present invention, when the image to be inspected is determined to be defective, the position of the defect is recorded, so the inspection result can be fed back to the previous process by referring to the recorded defect position. it can.

  The image inspection apparatus and the image inspection method of the present invention are not limited to the above-described embodiment, and can be variously modified without departing from the scope of the claims. For example, in the Mahalanobis Taguchi method, there is a method called item selection, and this method may be applied to the present invention to select items that contribute to inspection accuracy. Thereby, the inspection accuracy can be further improved.

  This item selection is to select an item that contributes to the inspection accuracy based on the SN ratio (the degree that the function does not vary), which is the concept of quality engineering. The basic procedure for selecting items is as follows.

1) Assign a plurality of pre-defined items to the orthogonal table as Level 1: Use item, Level 2: Do not use item.
2) A feature amount is extracted from the inspected image based on each condition of the orthogonal table.
3) The Mahalanobis distance of each condition of the orthogonal table is calculated from the extracted feature quantity.
4) The SN ratio of each item is calculated from the calculated Mahalanobis distance.
5) The item whose level 1 SN ratio is larger than the level 2 SN ratio is selected as an item contributing to the inspection accuracy. Further items are selected from the selected items as required (according to the inspection specification).

DESCRIPTION OF SYMBOLS 1 Image inspection apparatus 2 Camera 3 Illumination 4 Work 5 Electronic computer 6 Display apparatus 41 Good product image 411 Good product division image 42, 43 Test image 421, 431 Test division image 70 Defect S31 Good product image acquisition step S32 Good product image processing step S33 Good product Image segmentation step S34 First extraction step S35 First normalization step, first matrix calculation step, first distance calculation step S34, S35 Reference space creation step S31-S35 Reference space acquisition step, reference space acquisition means S36 Reference value setting step S41 Inspection image acquisition step, inspection image acquisition means S42 Inspection image processing step S43 Inspection image division step, inspection image division means S44 Second extraction step S45 Second normalization step, second Distance calculation step S44 S45 signal space creating step, the signal space creating means S46~S48 quality determining step, quality determination means S49 defect position recording step

Claims (12)

An image inspection method for determining the quality of an image,
A reference space acquisition step of acquiring a reference space composed of Mahalanobis distances for each of the good product divided images when the good product image is divided into a plurality of images (hereinafter referred to as “good product split images”);
Inspected image acquisition step of acquiring the inspected image;
An inspected image dividing step for dividing the inspected image into a plurality of images (hereinafter referred to as “inspected divided images”);
A signal space creating step for creating a signal space consisting of Mahalanobis distance for each of the inspected divided images based on the feature amount of the inspected divided image and the feature amount of the non-defective divided image;
By comparing the Mahalanobis distance of the corresponding signal space for each of the divided images to be inspected with a pass / fail judgment reference value determined based on the Mahalanobis distance of the reference space in the good product divided image at the same position. And a pass / fail determination step for determining pass / fail of the image.   The image inspection method according to claim 1, further comprising a defect position recording step of recording a position on the inspection image of the inspection divided image determined to be defective in the pass / fail determination step. The reference space acquisition step includes
A non-defective image acquiring step for acquiring the non-defective image;
A non-defective image dividing step for dividing the non-defective image into the non-defective divided images;
The image inspection method according to claim 1, further comprising: a reference space creation step of creating the reference space by calculating a Mahalanobis distance for each good product divided image. The non-defective image acquisition step acquires a plurality of non-defective images,
The reference space creation step includes
A first extraction step of extracting a plurality of predetermined feature amounts for each non-defective product divided image;
A first normalization step for normalizing those feature values and calculating good product standard feature values;
A first matrix calculating step of calculating a non-defective product correlation coefficient matrix comprising correlation coefficients between the non-defective product standard feature values for each of the non-defective product divided images at the same position between the plurality of non-defective products images;
And a first distance calculating step of calculating a Mahalanobis distance of the reference space based on the non-defective standard feature value and the non-defective correlation coefficient matrix for each non-defective divided image. 4. The image inspection method according to 3.   Including a reference value setting step for setting the pass / fail judgment reference value for each of the non-defective product divided images based on the maximum value of the Mahalanobis distances of the good product split images at the same position among the plurality of good product images. The image inspection method according to claim 4, wherein: The signal space creating step includes
A second extraction step for extracting the same feature quantity as the feature quantity extracted in the first extraction step for each of the inspected divided images;
A second normalization step of normalizing the feature amount extracted in the second extraction step with reference to the feature amount of the non-defective product divided image to calculate an inspection standard feature amount;
6. The second distance calculating step of calculating a Mahalanobis distance of the signal space based on the inspection criterion feature quantity and the non-defective product correlation coefficient matrix. 6. Image inspection method.   In the first and second extraction steps, when a specific horizontal line is drawn with respect to the luminance waveform of each image from which the feature amount is acquired, a differential value that is the number of intersections between the horizontal line and the luminance waveform, The image according to claim 6, wherein at least one of an integral value and a difference between a luminance maximum value and a luminance minimum value that are sections of the luminance waveform above a horizontal line is extracted as the feature amount. Inspection method.   The image inspection method according to claim 7, wherein the position of the horizontal line is variable. Inspected image processing step of performing predetermined image processing on the inspected image in order to easily detect defects in the inspected image;
The image inspection method according to claim 1, further comprising: a non-defective image processing step that performs the same image processing as the image processing on the non-defective image.   The image inspection method according to claim 3, wherein the non-defective image acquisition step is a step of acquiring an image included in an image group having the highest appearance frequency as the non-defective image. The inspected image is an image showing an appearance of a product to be inspected,
The image inspection method according to any one of claims 1 to 10, wherein the appearance of the product is an appearance in which the feature amount fluctuates greatly to the extent that the feature amount of the assumed defect is hidden. An image inspection apparatus that performs image quality determination,
Reference space acquisition means for acquiring a reference space consisting of the Mahalanobis distance for each good product divided image when the good product image is divided into a plurality of images (hereinafter referred to as “good product divided images”);
Inspected image acquisition means for acquiring the inspected image;
Inspected image dividing means for dividing the inspected image into a plurality of images (hereinafter referred to as “inspected divided images”);
Based on the feature amount of the inspected divided image and the feature amount of the non-defective divided image, signal space creating means for creating a signal space consisting of the Mahalanobis distance for each of the inspected divided images;
By comparing the Mahalanobis distance of the corresponding signal space for each of the inspected divided images with a pass / fail criterion value determined based on the Mahalanobis distance of the reference space in the good product divided image at the same position. An image inspection apparatus comprising: a quality determination unit that performs image quality determination.

Images