JP2005165482A - Defect detecting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To stably detect defects with little computational complexity even when there is little difference in the featured values of normal parts and defective parts, or even when the types of defects being the target of detection is variable when checking defects by image processing. <P>SOLUTION: A plurality of featured values p<SB>1</SB>to p<SB>m</SB>are extracted from an acquired image for an object to be tested, and a feature extraction image is prepared from each feature value, and a feature extraction array F<SB>e</SB>having those plurality of feature extraction images as components is created. The respective elements of the feature extraction array F<SB>e</SB>are plotted in a feature space having frequency information configured of a plurality of feature values, and a frequency image g<SP>(t)</SP>(x, y) is created by replacing the respective elements of the feature extraction array F<SB>e</SB>with frequency information acquired from the feature space. In the created frequency image g<SP>(t)</SP>(x, y), the part whose frequency is low is detected as a defective part in the acquired image, and the part whose frequency is high is detected as a normal part in the acquired image. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a defect detection method using an image processing technique.

  Although visual inspection has been conventionally performed for defect inspection of objects, it is a simple operation that requires continuous tension, and there are individual differences and physical conditions of inspectors, fluctuations in inspection standards, oversight, etc. There have been many attempts to automate defect inspection by image processing.

  In the inspection of defects by these image processing, a method for detecting defects by comparing the inspection image with the inspection object having no defect and examining the difference in the feature amount at the same image position is a printed circuit board. It has been adopted early in the field of inspection.

  On the other hand, in the case of defect detection in which it is difficult to set a standard for comparison, the following two approaches have been mainly used. One is a method of examining features of known defects and extracting those having those features from an inspection image. However, when there are various sizes and types of defects to be inspected, there is a disadvantage that inspection takes time.

  The other is a method of analyzing the characteristics of the normal part, examining the difference between the individual feature values, and detecting those that do not have the feature as defects, for example, obtaining the frequency domain range of the normal part, There is a method of comparing the detected image with the frequency domain. Such a method has an advantage that it can cope with a case where the type of defect is unknown, but has a disadvantage that a defect having a feature similar to the feature of the selected normal portion cannot be detected.

Specific literature describing the prior art will be described.
A method for comparing a test object without a defect with the inspection image and examining a difference in the feature amount at the same image position to detect a defect (compare with a two-dimensional image) is disclosed in Patent Document 1 and Patent It is described in Document 2.

  Among them, Patent Document 1 discloses normalized correlation matching in units of small regions in all corresponding images of a master image (reference) and a target image (inspection target) in order to detect a defect of a patterned industrial product. Is performed to extract a small area having a low matching rate and find a defect in the target image from the extracted small area.

  In Patent Document 2, in order to detect a defect on a metal surface, an inspection area on the surface of an object to be inspected is divided into a plurality of minute areas, and a standard deviation relating to luminance is measured in each of the minute areas. And the standard deviation in the masterwork without defects that have been measured and registered in advance, and by comparing this correlation with a preset threshold for detecting defects, A method for detecting the presence or absence of defects is described.

Methods for examining features of known defects and extracting those having those features from the inspection image are described in Patent Documents 3 to 6.
Among them, the method described in Patent Document 3 is a method for detecting only defects whose density change is relatively small, and the value obtained by the integration filter calculation processing in a predetermined block unit is used as the image data of the normal part. A defect on the object to be inspected is detected by comparison with the reference data based on it. However, density limiting means is provided in advance so that the density of the acquired image data falls within a predetermined range so as not to detect a defect having a large density change.

  In Patent Document 4, illumination having a black and white pattern is irradiated on an inspection surface, an acquired image is subjected to differential processing, binarization, expansion / contraction, and labeling processing are performed, and defects having a predetermined number of pixels or more are defective. Is described.

Patent Document 5 describes a method of comparing image data at each inspection point of an inspection object with a predetermined average value in order to detect an appearance defect of the inspection object.
In Patent Document 6, in order to detect contamination on the surface of a magnetic head or a magnetic disk, a difference between an average of adjacent pixels and an attention point is obtained from an acquired image, and the difference value is compared with a reference value. Describes a method for detecting a defect of the inspection object.

  As a method of analyzing the features of the normal part, examining the differences between the individual feature amounts, and detecting those having no features as defects, for example, the range of the frequency region of the normal part is obtained, and the detected image and frequency There are methods to compare by area (comparison by feature space).

  For example, in Patent Document 7, in order to detect defects occurring in a periodic pattern, the inspection area is an integral multiple of the vertical and horizontal periods of the periodic pattern without defining a correlation with the periodic pattern. A method is described in which an average of a plurality of divided rectangular areas is obtained by dividing into rectangles of the same size, a ratio between the average and each rectangular area is obtained, and the presence or absence of a defect in the rectangular area is determined from the ratio. Has been.

Non-Patent Document 1 quantifies the level of abnormalities allowed for non-defective products in the frequency domain of the image of the inspection object using a standard statistical method, and compares the value range with the frequency domain value of the inspection object Thus, a method for detecting an out-of-range object as a defect is described.
JP 2002-314882 A JP 2001-126064 A JP 2002-230522 A Japanese Patent Laid-Open No. 2002-5844 JP 2001-175859 A JP-A-8-235543 JP 2001-281166 A Satoshi Nakayama, Ryuji Sakita, Terumi Kamada: Development of visual inspection method by frequency analysis, Proc. Of the 8th Image Sensing Symposium, pp.145-148, July 2002

  Defects where the normal and defective areas are unclear and poor in contrast are difficult to detect by local processing such as edge detection, and the feature amount is almost the same as the normal area. It is difficult to detect even with the method. In addition, the technique of examining the characteristics of known defects and extracting those having these characteristics from the inspection image has a drawback that inspection takes time if the sizes and types of defects to be inspected are various.

  Therefore, the present invention can detect a defect stably and with a small amount of calculation even if there is little difference in the feature quantity between the normal part and the defective part, and the types of defects to be detected are various. The purpose is to do so.

  In order to achieve this object, the present invention plots a value determined by a combination of a plurality of feature amounts extracted at each pixel position of an acquired image of an inspection object in a feature space composed of the feature amounts, Frequency information is given to each point plotted on the feature space, and the frequency of the value determined by the combination of the feature quantities is obtained from the feature space at each pixel position of the acquired image. From this frequency, the frequency is low A part is detected as a defective part, and a part with high frequency is detected as a normal part.

  The present invention also extracts a plurality of feature amounts from the acquired image of the inspection object, creates a feature extraction image from each feature amount, and creates a feature extraction array having these plurality of feature extraction images as components. Plotting each element of the feature extraction array in a feature space having frequency information composed of the plurality of feature quantities, and examining the frequency of each element of the feature extraction array and the corresponding position in the feature space The extracted low-frequency part is detected as a defective part in the acquired image, and the high-frequency part is detected as a normal part in the acquired image.

  Specifically, the defect is regarded as a part having some feature quantity different from a normal part other than the defect, and first, m feature quantities are extracted from the image obtained by photographing the inspection area as a natural number, and each extracted feature quantity is obtained. And a feature extraction array having these feature extraction images as components. That is, each element of the feature extraction array is a vector based on the extracted m feature amounts (this vector is referred to as a “feature vector”), and represents a feature at each point of the original acquired image. After that, an m-dimensional feature space having frequency information at each coordinate is constructed based on this feature extraction array, and all elements of the feature extraction array are plotted in this feature space, and then each element of the feature extraction array appears. The frequency is checked, and the low-frequency portion is detected as a defective portion in the acquired image, and the high-frequency portion is detected as a normal portion in the acquired image.

  In this case, since the appearance frequency of the pixel having the feature of the defective portion is smaller than that of the pixel having the feature of the normal portion, it is difficult to distinguish even if the difference in the feature amount between the defective portion and the normal portion is small. Even a defect can be easily detected.

  Further, the present invention extracts a plurality of feature amounts from an acquired image of an object, creates a feature extraction image from each feature amount, creates a feature extraction array having these plurality of feature extraction images as components, A frequency image in which each element of the feature extraction array is plotted in a feature space having frequency information composed of the plurality of feature amounts, and each element of the feature extraction array is replaced with frequency information obtained from the feature space. By creating, the frequency and position of the value determined by the combination of a plurality of feature amounts extracted from the acquired image are simultaneously acquired.

  That is, m feature amounts are extracted from the acquired image, a feature extraction image is created from each feature amount, a feature extraction array having these m feature extraction images as components is created, and m feature features are extracted. After plotting each element of the feature extraction array in an m-dimensional feature space with frequency information composed of quantities, create a frequency image by replacing each element of the feature extraction array with the frequency information obtained from the feature space To do.

  In this way, at each point of the acquired image, a feature extraction array expressing the features of each point as a feature vector by combining m feature quantities is created, and each element of this feature extraction array is extracted from the feature space. Since a frequency image is created by replacing with the obtained frequency, a new image can be obtained by replacing the feature of each point of the original image with the appearance frequency of the feature, and the appearance frequency of similar features can be imaged Can do.

  Furthermore, the present invention extracts a plurality of feature amounts from an acquired image of an inspection object, creates a feature extraction image from each feature amount, and creates a feature extraction array having these plurality of feature extraction images as components. A frequency image in which each element of the feature extraction array is plotted in a feature space having frequency information composed of the plurality of feature amounts, and each element of the feature extraction array is replaced with frequency information obtained from the feature space In the generated frequency image, a low-frequency portion is detected as a defective portion in the acquired image, and a high-frequency portion is detected as a normal portion in the acquired image.

  In this way, when the frequency image is created by the above procedure from the image acquired in the defect inspection, the low-frequency part of the frequency image is detected as a defective part in the defect inspection image, and the high-frequency part is detected. It can be detected as a normal part in the defect inspection image.

  Further, the present invention creates a frequency image representing the appearance frequency of values determined by a combination of a plurality of feature amounts extracted from each image from a plurality of acquired images of the inspection object, and from these plurality of frequency images A total image or an average image of these frequency images is obtained, and a defect range is detected by designating a frequency range to be detected from the total image or the average image.

That is, a frequency image representing an appearance frequency of a value determined by a combination of m feature amounts extracted from each image is created from a plurality of acquired images, and a total image of the frequency images is generated from the plurality of frequency images. Alternatively, an average image is obtained, and a defective range is detected by designating a frequency range to be detected from the total image or the average image.
If it does in this way, the defective part in a defect inspection image can be detected easily.

  In the above description, the feature extraction image and the feature extraction array have been described as having the same size as the original image. However, a feature extraction image and a feature extraction array having a size larger or smaller than the original image and a frequency image are created. May be. However, it is necessary to keep correspondence between the pixel of the feature extraction image having the changed size, the element of the feature extraction array, the pixel of the frequency image, and the pixel of the original image.

  According to the present invention, as described above, it is possible to detect a defect with a poor boundary and poor contrast, which is difficult to understand only by human observation. In addition, by changing the detection parameter, it is possible to separate and detect a clear defect and a defect with low contrast. Furthermore, by creating a frequency image, the feature of the defect can be visualized, and the size and position of a plurality of defects can be detected simultaneously.

  In addition, according to the present invention, since frequency images obtained from a plurality of images are added, noise can be removed and image processing can be performed quickly with a small amount of calculation.

  In the field of pattern recognition, feature spaces are often used for the purpose of classifying observed patterns into predetermined classes. That is, in pattern recognition, a plurality of essential features effective for classification are extracted from an original pattern, each feature is quantified, and a feature space that combines them is used. Now, if m features are used with m as a natural number, the pattern is represented as one point in an m-dimensional feature space.

  In the present invention, m features are extracted from an acquired image obtained by photographing an inspection region, and feature extraction images are created by imaging each of the extracted m feature amounts, and then feature extraction is performed by collecting a plurality of them. Create an array.

That is, the feature extraction array can be said to be an array in which each element of the array is an m-dimensional vector having m feature values as components.
A feature vector p composed of m feature quantities is expressed as follows using m feature quantities p _{1 to} p _m .

p = (p ₁ , p ₂ ,..., p _m ) ^t (1)
Here, t represents transposition.
When a plurality of feature quantities are extracted at each point of the acquired image, the coordinate value of each point of the acquired image is (x, y), and the feature vector corresponding to the coordinates is p _{x, y} , _{x in} the feature extraction array The elements in row y are px _{, y} .

  The feature quantity extracted from the acquired image is the feature of the point (density value, red component, etc.) determined only by the individual pixels of the acquired image, and the feature (line, edge strength, Edge direction, average value, etc.), features (shape, connectivity, etc.) within a certain region can be used.

The feature extraction images of the feature quantities p ₁ , p ₂ ,..., P _m are respectively represented as f _e1 , _{fe 2} ,..., F _em , and the feature extraction array is _Fe , and FIG. A creation example is shown.
Now, when m-dimensional feature space is constructed using m feature quantities effective for defect detection, and each element of the feature extraction array is plotted in the feature space, the normal part becomes the feature as shown in FIG. Concentrate on a certain area in the space. On the other hand, defective portions are scattered in other regions.

  Assuming that each point plotted in the feature space has the frequency of that point, the frequency of each point in the normal part region is high, and the frequency of each point in the defective part is low. In FIG. 2, the points with higher frequency are displayed brighter.

  Defects that are difficult to distinguish often have a small difference in feature quantity from other parts, and are plotted near the area where the normal part is plotted on the feature space. However, the frequency is small compared to the normal part.

As described above, the normal portion and the defective portion can be identified by examining the frequency of each plotted point in the feature space having the frequency information obtained from the image including the defect.
However, in an actual scene, pixel values are often affected by various noises, and detection is performed using a plurality of images for the purpose of noise removal.
The procedure is shown below.

(1) An image is represented by a function f (x, y) of two independent variables x and y representing the coordinates. After extracting m feature values from the image f (x, y) of the inspection area, create a two-dimensional array of the same size as this image, and set the value of each element of the array to f (x, y) A feature extraction array F _e (x, y) set to an m-dimensional feature vector corresponding to the coordinate value of is generated.

(2) a feature extraction sequence acquired at time t and ^{_{F e (t) (x,}} y), F e (t) (x, y) the F _e ^(t) (x, y) obtained from m Let h (t) be plotted on a feature space composed of feature quantities. h (t) has the coordinate component of each feature quantity in the feature space and the frequency at that coordinate.

(3) When the feature vector of each element of the original feature extraction array F _e ^(t) (x, y) is replaced with the frequency obtained from h (t), F _e ^(t) (x, y) becomes It becomes a grayscale image representing the appearance frequency of feature vectors. This image is called a frequency image g ^(t) (x, y).

(4) A total S (x, y) of n frequency images g ^(t) (x, y) obtained from {F _e ^(t) (x, y), t = 1, 2,. Obtained by the following equation.

以上の手順は図３に示されている。

The above procedure is shown in FIG.

Furthermore, if the average of the frequency images is g _ave (x, y), g _ave (x, y) is expressed as follows.

g_ave(x、y)を平均頻度画像と呼び、g_ave(x、y)の各画素の値は実数値を取るものとする。

g _ave (x, y) is called an average frequency image, and the value of each pixel of g _ave (x, y) is assumed to be a real value.

(5) The frequency threshold value for identifying the defective portion and the normal portion is d, and this is called the detection average frequency. Since the defective portion has a lower frequency than the normal portion, a portion having a value smaller than the threshold value d among the pixels of g _ave (x, y) is detected as a defective portion. When there are a plurality of defect portions, the type and size of the detected defect vary depending on the value of d.
Detection is performed by the following equation.

検出されたg_b(x，y)は２値画像で、g_b(x，y)の0-画素の部分が欠陥部分を表す。

The detected g _b (x, y) is a binary image, and the 0-pixel portion of g _b (x, y) represents a defective portion.

(6) When there are a plurality of defective portions, it is easier to understand if the pixel values equal to or lower than the threshold value d are displayed as grayscale images instead of binary images.

  Using an i-bit image memory, a frequency image up to the detection average frequency d is expressed by the following equation.

ここで、[ ]はガウス記号を表す。

Here, [] represents a Gaussian symbol.

The detected gray image g _d (x, y) is called a frequency feature detection image. The dark part of g _d (x, y) represents a defect. A bright part of g _d (x, y) represents a defect whose feature is close to a normal part or a large total area of the defective part.

(Example 1)
Next, specific examples of the present invention will be described. Here, the following experiment was conducted.

  That is, in the experiment, first, the minute dust adhering to the surface of the CD-R case in which thick paper with ruled lines was accommodated as shown in FIG. 4 was detected. The detection was performed on a part of the CD-R case as shown in FIG. In the case of an inspection object having a transparent film on the surface, dust adhering to the surface may be erroneously detected as an internal defect and cause a problem.

  Such surface dust can be easily detected by using a light source incident at a shallow angle with respect to the inspection surface. However, with a small illumination device incorporated in a general inspection apparatus, it is difficult to uniformly irradiate the inspection surface at a shallow angle, and normally, the side closer to the light source is brighter, and the light is irradiated darker as the distance increases. In addition, some dust is not detected depending on the irradiation direction of the light source, and it is desirable to examine by irradiating from multiple directions.

  In the experiment, the CD-R case was placed horizontally. Then, using red, green and blue LED bar illumination, the inspection surface was irradiated at a shallow angle from three directions, and this was photographed with a 3CCD camera from directly above. The RGB components of the image at this time are shown in FIGS.

  FIGS. 9 and 10 show the result of creating a feature extraction array using the pixel values of the image acquired in the direction of each light source as a feature amount and detecting surface dust by this method. FIG. 9 shows d = 3, n = 5, and i = 8 in the above-described equations, and only the dust on the surface of the CD-R case can be detected black. FIG. 10 shows the case where d = 10, n = 5, and i = 8 in the above-described equations. By changing the detection threshold d, even the ruled line printed on the paper inside the CD-R case is detected. doing.

  FIG. 11 shows a result of applying a publicly known Sobel filter in the horizontal and vertical directions to the five addition average images of each image of FIGS. Here, dust having a clear boundary can be detected, but dust having a clear boundary is difficult to distinguish from ambient noise. Further, in FIG. 11, since the boundary of dust is detected by the Sobel filter, each dust is detected larger than the actual size.

(Example 2)
Next, a defect detection experiment with poor contrast was performed.

  Characters were written with a ballpoint pen on the first sheet of ruled memo paper (Kokuyo Me-91) as shown in FIG. 12, and the handwriting left behind by the pressure on the third memo paper was photographed. The same illumination as in the previous experiment was used, and illumination was performed at a shallower angle with respect to the inspection surface than the three directions as in the previous experiment. An image taken with a 3CCD camera is shown in FIG.

  By illuminating the illumination at a shallow angle from a slant, the surface irregularities can be emphasized, but in the case of this inspection object, even fine irregularities due to fibers on the paper surface were emphasized. Therefore, first, the “smoothing of 5 neighborhoods” processing is performed on each acquired image, and then the irradiation unevenness of each illumination is corrected. Specifically, an image obtained by performing “smoothing of 9 neighborhoods” on the original image five times is Z (x, y), and an average value of all the pixels of Z (x, y) is L. The feature extracted image is obtained by multiplying each pixel of the image after “smoothing of 5 neighborhoods” by L / Z (x, y). This feature was obtained for images in each irradiation direction, and a feature extraction array was created.

  The detection result is shown in FIG. It can be seen that the boundary is unclear and almost undistinguishable.

(Example 3)
Next, the defect of the bearing surface which is a metal component was detected. That is, an inner ring of a ball bearing having a stepped portion on the surface as shown in FIG. Examples of defects on the surface of the inner ring, that is, the ball bearing component include scratches that occur during conveyance of the component, scratches that occur on the grinding surface when the component cannot be smoothly chucked during processing, and adhesion of chips. When these are photographed with a camera, the wound is photographed brighter than a normal metal surface, and scratches, chips, etc. are photographed darkly, and their appearance changes considerably depending on the illumination angle of illumination.

  The ring illumination corresponding to the shape of the inner ring was used, and the irradiation angle was changed by changing the height in three stages to acquire an image of the inspection target part. The images are shown in FIGS. FIGS. 19 to 21 show the result of detecting the original image by compressing the 8-bit image data to 5 bits to create a feature extraction array and changing the value of the detection parameter d together with a partially enlarged view at the bottom of the image. . FIG. 19 shows d = 10, n = 5, and i = 8 in each of the above equations, and FIG. 20 shows d = 200, n = 5, and i = 8 in each of the above equations. FIG. 21 shows d = 600, n = 5, and i = 8 in the above-described equations.

  The inspection surface of the bearing as the metal part has minute scratches such as polished eyes, and gradation compression is performed to prevent overdetection of these minute scratches. Looking at the detection results, fine polishing eyes are detected in FIG. 21 where the value of d is large, and scratches that strongly reflect light are detected in FIG. 19 where the value of d is small.

  Thus, the defect which can be detected changes by changing a detection parameter, and it is dependent on the size of a defect, and the difference in the distance in the feature space of a defect part and a normal part. That is, distinct defects are detected with a small d value, and unclear defects are detected with a large d value.

It is a figure which shows the creation example of the feature extraction arrangement | sequence based on this invention. It is a figure which illustrates distribution of the normal part and defect part in feature space based on this invention. It is a figure which shows the preparation procedure of the frequency image using the feature space from the feature extraction arrangement | sequence, and the sum total of a frequency image based on this invention. It is a figure which shows the CD-R case used for the experiment based on this invention. FIG. 5 is an enlarged view showing a portion targeted for image processing in the CD-R case of FIG. 4. It is a figure which shows the example of the red component image of the image which image | photographed the part of FIG. It is a figure which shows the example of the green component image of the image which image | photographed the part of FIG. It is a figure which shows the example of the blue component image of the image which image | photographed the part of FIG. It is a figure which shows an example of the detection result of the surface dust about the image which image | photographed the part of FIG. It is a figure which shows the other example of the detection result of the surface dust about the image which image | photographed the part of FIG. It is a figure which shows the example which applied the well-known Sobel filter with respect to the image which image | photographed the part of FIG. It is a figure which shows the example of the image which image | photographed the character written with the ball-point pen on the 1st sheet of memo paper. It is a figure which shows the example of the image which image | photographed the handwriting left by the pen pressure on the 3rd sheet of the memo paper. It is a figure which shows the example of the frequency characteristic detection image detected based on this invention regarding the image of FIG. It is a perspective view of the bearing inner ring | wheel which is a metal component as a measuring object. It is a figure which shows the example of the image which image | photographed the surface of the bearing of FIG. 15 by the illumination in a high position. It is a figure which shows the example of the image which image | photographed the surface of the bearing of FIG. 15 by the illumination in an inside position. It is a figure which shows the example of the image which image | photographed the surface of the bearing of FIG. 15 by the illumination in a low position. It is a figure which shows the example of the frequency feature detection image detected based on the image of FIGS. It is a figure which shows the example of the frequency feature detection image which detected by changing the parameter for a detection compared with FIG. It is a figure which shows the example of the frequency characteristic detection image which detected by changing the parameter for a detection compared with FIG.

Claims (5)

  A value determined by a combination of a plurality of feature amounts extracted at each pixel position of the acquired image for the inspection object is plotted in a feature space composed of the feature amounts, and each point plotted on the feature space Frequency information is obtained, and the frequency of the value determined by the combination of the feature quantities is obtained from the feature space at each pixel position of the acquired image, and a low-frequency portion is detected as a defective portion from this frequency. A defect detection method comprising detecting a high part as a normal part.   A plurality of feature amounts are extracted from the acquired image of the inspection object, a feature extraction image is created from each feature amount, a feature extraction array having these plurality of feature extraction images as components is created, and the plurality of features Plot each element of the feature extraction array in a feature space having frequency information composed of a quantity, extract the frequency of each element of the feature extraction array by examining the corresponding position in the feature space, A defect detection method comprising: detecting an extracted low-frequency part as a defective part in the acquired image, and detecting a high-frequency part as a normal part in the acquired image.   A plurality of feature amounts are extracted from the acquired image of the object, a feature extraction image is created from each feature amount, a feature extraction array having these plurality of feature extraction images as components is created, and the plurality of feature amounts By plotting each element of the feature extraction array in a feature space having frequency information constituted by creating a frequency image in which each element of the feature extraction array is replaced with frequency information obtained from the feature space, An image processing method, wherein a frequency and a position of a value determined by a combination of a plurality of feature amounts extracted from the acquired image are acquired simultaneously.   A plurality of feature amounts are extracted from the acquired image of the inspection object, a feature extraction image is created from each feature amount, a feature extraction array having these plurality of feature extraction images as components is created, and the plurality of features Create a frequency image by plotting each element of the feature extraction array in a feature space having frequency information composed of quantities, and replacing each element of the feature extraction array with frequency information obtained from the feature space A defect detection method, comprising: detecting a low-frequency portion as a defective portion in an acquired image and detecting a high-frequency portion as a normal portion in the acquired image in the obtained frequency image.   A frequency image representing the appearance frequency of a value determined by a combination of a plurality of feature amounts extracted from each image is created from a plurality of acquired images of the inspection object, and these frequency images are generated from the plurality of frequency images. 5. The defect detection method according to claim 4, wherein a defect portion is detected by obtaining a total image or an average image, designating a range of frequencies to be detected from the total image or the average image.

Images