JP2011089920A - X-ray inspection method and x-ray inspection apparatus using the same - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an X-ray inspection method and an X-ray inspection apparatus highly reliable by providing appropriate determination conditions on the basis of learning on a distribution of the amount of characteristics of normal products. <P>SOLUTION: In the X-ray inspection method, measurements, inspection, and the generation of characteristic vectors are performed on a plurality of normal products prior to inspection on an object to be inspected (Step S1-S7). A plurality of acquired characteristic vectors are stored as learning data (Step S8-S10) to generate a normal product region (Step S11). Measurements, inspection, and the generation of characteristic vectors are similarly performed on the object to be inspected (Step S12-S18) to determine the quality of the object to be inspected on the basis of whether its characteristic vectors are included in the normal product region or not (Step S19). Since the normal product region generated in the Step S11 is acquired from a distribution of normal products, the determination in the Step S19 includes the correlation between amounts of characteristics. It is thereby possible to perform highly reliable determination. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention uses a plurality of types of feature amounts obtained from an X-ray image subjected to image processing on an article to be inspected (hereinafter referred to as “inspected object”), and detects foreign matter contamination and defects. The present invention relates to an X-ray inspection apparatus that performs various inspections such as inspection and number inspection.

  An X-ray inspection device is a device that non-destructively inspects the interior of an article by irradiating the article with X-rays, measuring the amount of X-ray transmission with an X-ray detector, and creating an X-ray image based on the X-ray detector. is there. For example, foreign matter contamination inspection for inspecting whether foreign matter is not mixed in the contents of the packaged product, number inspection for inspecting whether the content has a predetermined number, cracks in the content, etc. There is a defect inspection for inspecting whether or not a defect has occurred. These inspections are performed on various articles such as food, pharmaceuticals, and industrial parts.

  X-ray inspection apparatuses are roughly classified into two types: X-ray fluoroscopy apparatuses and X-ray CT (Computed Tomography) apparatuses. Among these, the X-ray fluoroscopic apparatus irradiates X-rays from one direction to obtain a transmission image, whereas the X-ray CT apparatus performs Fourier transform on the data of the X-ray transmission amount measured in the entire circumference direction of the inspection object. A tomographic image of the internal structure is obtained by reconstruction by conversion.

  Since the X-ray fluoroscopic apparatus only performs fluoroscopy from one direction, the X-ray fluoroscopic apparatus has a feature that although the amount of information is smaller than that of the X-ray CT apparatus, the inspection object can be inspected quickly. Therefore, it is frequently used for continuous inspection of articles continuously conveyed by a belt conveyor or the like. On the other hand, the X-ray CT apparatus requires imaging from the entire circumference and reconstruction of a tomographic image, and requires more processing than the X-ray fluoroscopic apparatus. Although it is possible to process this at high speed, it is necessary to improve the shooting speed, the processing speed of the computer that performs image reconstruction, and the like, which increases the manufacturing cost of the apparatus. Therefore, it is more suitable for precision inspection than continuous inspection that requires high-speed processing.

  In these X-ray inspection apparatuses, predetermined image processing is performed on the X-ray image of the object to be inspected, one or a plurality of feature amounts are extracted therefrom, and a reference value (allowable range) given in advance by learning or the like is extracted. ) To determine whether the object to be inspected is a normal product. As the above-described image processing, a binarization process for dividing the acquired X-ray image into a background image area and an image area other than the background, and a labeling for classifying the image area divided by the binarization process There are processing. The extracted feature amount includes the perimeter and area of the image region (Patent Document 1). Furthermore, from the X-ray transmission amount in the image area, thickness, volume, and the like can be extracted as a feature amount (Patent Document 2).

Japanese Unexamined Patent Publication No. 2005-031069 JP 2007-183200 A

  As can be seen in the X-ray inspection apparatuses of Patent Document 1 and Patent Document 2, in the inspection method in the conventional X-ray inspection apparatus, it is determined independently whether each feature amount is within a corresponding allowable range. As a result, the quality of the inspection object is judged. However, in some cases, products that are determined to be normal products through this inspection method are determined as defective products in the final check by visual inspection or precise inspection after that.

  Such a problem can easily occur when each feature amount is in the vicinity of the limit of the allowable range. For example, in the case of an inspection in which parameters of the length and width of an inspection object are acquired as feature quantities, if both feature quantities exist near the limit of the allowable range, the area of the inspection object is within the range of a normal product. Sometimes it deviates. In recent years, consumer demand for the management of foods and the like has become stricter, and accordingly, the detection accuracy of foreign matters, defective products, and the like required for X-ray inspection apparatuses has increased. Therefore, in order to prevent such a misjudgment, in the above example, the type of feature quantity is increased, for example, the area parameter is added to the feature quantity, or the allowable range for the length and width feature quantities is set more strictly. There is a need.

  However, when the types of feature quantities are increased, processing for extracting them increases, and as a result, processing time and device manufacturing costs increase. In addition, if the allowable range is set strictly, an object to be inspected that should be included in a normal product is determined as a defective product, and yield rate and work efficiency deteriorate.

  The problem to be solved by the present invention is to provide a highly reliable X-ray inspection method and X-ray inspection apparatus that solve the above-mentioned problems by providing appropriate determination conditions.

An X-ray inspection method according to the present invention made to solve the above problems is as follows:
A step of generating an X-ray image of the article by irradiating the article with X-rays and measuring the amount of transmission, and extracting n types (n is an integer of 2 or more) of feature quantities from the X-ray image And a step of generating an n-dimensional feature vector based on the extracted n types of feature quantities,
generating feature vectors for m (m is an integer of 2 or more) normal products, and storing the distribution of feature vectors of the normal products as learning data;
Determining a normal product region based on the learning data;
Generating a feature vector for the inspection object, and determining whether the inspection object is a normal product according to whether the feature vector of the inspection object is included in the normal product region;
It is characterized by having.

Further, an X-ray inspection apparatus according to the present invention made to solve the above problems is
Image generation means for generating an X-ray image of the article by irradiating the article with X-rays and measuring the amount of transmission, and n types (n is an integer of 2 or more) of feature quantities from the X-ray image In an X-ray inspection apparatus comprising: a feature quantity extracting unit that extracts a feature vector; and a feature vector generating unit that generates an n-dimensional feature vector based on the n types of extracted feature quantities.
learning means for generating feature vectors for m (m is an integer of 2 or more) normal products, and storing the distribution of the feature vectors of the normal products as learning data;
Based on the learning data, a region determining means for determining a normal product region;
Determining means for generating a feature vector for the object to be inspected, and determining whether the object to be inspected is a normal product according to whether the feature vector of the object to be inspected is included in the normal product region; ,
It is characterized by having.

  The space in which the feature vector exists may be a space having each feature parameter as a coordinate axis (hereinafter referred to as “feature space”) or a mapping space thereof. The function (mapping function) for converting from the feature space to the mapping space may be either linear or non-linear.

  In addition, various methods such as a support vector machine, naive Bayes classifier, machine learning such as cluster analysis, and a technique in the data mining field can be used for determining the normal product region.

  In the present invention, a normal product region is determined based on a distribution of feature vectors acquired in advance for a plurality of normal products, and then the device to be inspected depends on whether or not the feature vector of the device to be inspected is included in the normal product region. Judge the quality of things. Since the normal product region includes the correlation between the feature amounts, the determination using the normal product region necessarily includes the correlation between the feature amounts. Thereby, it is possible to appropriately determine the inspection object, and it is possible to provide an X-ray inspection method and an X-ray inspection apparatus that are more reliable than those in the past. Further, since the normal product region can be automatically set from the distribution of the feature vector of the normal product, it is not necessary for the user to enter a numerical value within the allowable range, and the usability can be improved.

The graph which shows the example of the defect which cannot be detected with the conventional X-ray inspection method. The block diagram which shows the structure of one Example of the X-ray detection apparatus which concerns on this invention. The flowchart which shows the procedure of operation | movement of the X-ray detection apparatus of a present Example. The schematic diagram which shows the procedure of the production | generation of the normal product area | region in the X-ray detection method of a present Example.

  Since conventional X-ray inspection methods only combine independent inspections for each feature quantity, if there is a correlation between multiple feature quantities, defective products that are out of the normal product distribution are also normal products. May be considered. This will be described with reference to FIG.

  As shown in FIG. 1A, in the inspection using, for example, the length and the width as the feature quantity, the area of the normal product may have to be within a certain range. In such a case, if both the length and width are close to the limit of the larger allowable range, or vice versa, even if each feature amount is within the allowable range, the area is reduced. It will deviate from the area where normal products are distributed. Further, as shown in FIG. 1B, in the inspection using the area and the volume as the feature amount, the ratio of these feature amounts may have to be within a certain range due to problems such as shape. Therefore, when a plurality of feature quantities have some relationship with each other, it is necessary to determine the quality of the inspection object in consideration of the correlation.

  Therefore, in the X-ray inspection apparatus according to the present invention, after generating a normal product region including a distribution of feature vectors of a plurality of normal products using the feature vectors of a plurality of normal products as the learning data, the feature vector of the inspection object is normal. The quality of the object to be inspected is determined depending on whether it is included in the product area. Since the normal product region is obtained from the distribution of normal products, it necessarily includes a correlation between feature quantities. Therefore, even when the conventional inspection method as shown in FIGS. 1 and 2 is determined as a normal product despite being out of the distribution of the normal product, it can be appropriately determined as a defective product. .

  An embodiment of the X-ray inspection apparatus according to the present invention will be described with reference to FIGS. Here, FIG. 2 is a block diagram showing the configuration of the X-ray inspection apparatus of the present embodiment, FIG. 3 is a flowchart showing the procedure of inspection by the X-ray inspection apparatus, and FIG. 4 is normal in the X-ray detection method of the present embodiment. It is a schematic diagram which shows the procedure of a production | generation area | region.

The configuration of the X-ray inspection apparatus of this embodiment shown in FIG. 2 will be described below.
The X-ray inspection apparatus according to the present embodiment irradiates a normal product or an object to be inspected with X-rays and measures the amount of transmission, and the X-ray transmission based on the amount of X-ray transmission measured by the measurement unit 10. An image is created, n types (n is an integer of 2 or more) feature quantities are extracted, an inspection unit 20 that generates an n-dimensional feature vector, and a distribution of a plurality of normal product feature vectors are learned. A pre-processing unit 30 that determines an area and a determination unit 40 that determines pass / fail of an inspection object based on a normal product area determined by the pre-processing unit 30.

  The measurement unit 10 described above includes an X-ray irradiation unit 11 that emits X-rays from an X-ray source, and an X-ray detector that measures the X-ray transmission amount of a normal product or an inspection object irradiated with X-rays. And an X-ray detection unit 12 made up of, and the like. The inspection unit 20 is generated by the image generation unit 21 and the image generation unit 21 that generate X-ray images based on the X-ray transmission amounts of normal products and inspection objects measured by the X-ray detection unit 12. An image processing unit 22 that performs image processing on an X-ray image, a feature amount extraction unit 23 that extracts n types of feature amounts from the X-ray image subjected to image processing by the image processing unit 22, and a feature amount extraction unit 23 And a feature vector generation unit 24 that generates an n-dimensional feature vector from the extracted feature quantity. Further, the preprocessing unit 30 determines whether to send the feature vector data to the preprocessing unit 30 depending on whether the feature vector data generated by the feature vector generation unit 24 is a normal product or an inspection object. A switching unit 31 that switches whether to send to the unit 40, a learning unit 32 that learns and stores feature vectors of a plurality of normal products sent to the preprocessing unit 30 by the switching unit 31, and a normal product region from the distribution And an area determination unit 33 to be determined.

Next, the inspection procedure of the X-ray inspection apparatus of this embodiment shown in FIG. 2 will be described with reference to the flowchart of FIG.
In the X-ray inspection apparatus of the present embodiment, before performing inspection on the inspection object, each feature vector is generated for a plurality of normal products and stored as learning data (steps S1 to S10). The generation and learning of the feature vector is performed as follows.

  The switching unit 31 in the preprocessing unit 30 is operated so that the feature vector data obtained by the inspection unit 20 is sent to the preprocessing unit 30, and the measurement unit 10 is normal at a predetermined position (X-ray irradiation position). The product is sent (step S1). Then, the normal product is irradiated with X-rays from the X-ray irradiation unit 11 (step S2), and the amount of transmission is measured by the X-ray detection unit 12 (step S3). The X-ray transmission data measured by the X-ray detector 12 is sent to the image generator 21 in the inspection unit 20 to generate an X-ray image for the normal product (step S4). However, since the X-ray image generated by the image generation unit 21 includes not only a normal product image but also a background image, the image processing unit 22 performs predetermined image processing to obtain a background image and a normal product image. Must be separated. For this image processing, for example, binarization processing or labeling processing can be used. Thereby, the region of the target article (in this case, a normal product) can be detected from the X-ray image (step S5).

For the normal product image region detected by the image processing unit 22, the feature amount extraction unit 23 extracts n types of feature amounts x ₁ , x ₂ ,..., X _n used for inspection (step S6). . Using the n types of extracted feature quantities x ₁ , x ₂ , ..., x _n , feature vectors X = [x ₁ , x ₂ ,. .., x _n ] are generated by the feature vector generation unit 24 (step S7). As the feature amount used for the inspection, for example, a known image feature amount such as an area, a volume, a perimeter length, a maximum length in the inertial principal axis direction, and a circularity can be used. Furthermore, the feature vector data of the normal product generated by the feature vector generation unit 24 is sent to the learning unit 32 in the preprocessing unit 30 and stored as learning data (step S8).

By repeating the above processing for other normal products, feature vector learning data for a plurality of normal products can be obtained (steps S9 and S10). Here, the number of normal products that have generated feature vectors is m, the feature vectors of m normal products are X ₁ , X ₂ ,..., X _m , and the set is the learning data A = {X ₁ , X ₂ , ..., X _m }. It is desirable that the number m of normal products is at least as large as the number n of feature types used for inspection, and the larger the number, the better.

Next, a normal product region is determined from the learning data A obtained by the learning unit 32 (step S11). Hereinafter, the method for determining the normal product region used in this embodiment will be described with reference to FIG.
FIG. 4 (a) schematically shows the distribution of each element (normal feature vector) of the learning data A in the n-dimensional feature space. Here, it is assumed that these distributions can be approximated by a multidimensional normal distribution with an average X. Further, it is assumed that the normal feature vector is in a range in which the probability density function of the multidimensional normal distribution takes a value equal to or greater than a certain threshold value. Under these assumptions, whitening is performed after translation so that the average of _m normal product feature vectors X ₁ , X ₂ , ..., X _m included in learning data A is zero vector Then, the normal product region in the mapping space is a spherical region having a predetermined radius or less with the origin at the center. Hereinafter, this predetermined radius is represented by R.

とし、固有ベクトルを列ベクトルに持つ行列をE=(e₁ ^T e₂ ^T ... e_n ^T)とすると、集合A'の要素に対する白色化行列Vは、V = DE^Tで計算することができる。なお、白色化行列の計算方法は一通りでなく、別の計算方法によって得られた白色化行列を本発明に用いることもできる。例えば、ある白色化行列V'に任意の直交行列Uを掛けたUV'も白色化行列となる。前記の行列Eは直交行列であるため、EDE^Tも白色化行列となる。従って、DE^Tの代わりにEDE^Tを白色化行列Vとして用いても良い。 Specifically, if X _i '= X _i −X _a and A' = {X ₁ ', X ₂ ', ..., X _m '}, this set A' is the feature vector of the training data A. The average is translated so as to be a zero vector (FIG. 4B). Where the covariance matrix for the elements of set A 'is C, the eigenvalues of this covariance matrix C are d ₁ , d ₂ , ..., d _n , and the eigenvectors with norm _equal to 1 corresponding to each eigenvalue e ₁ , e ₂ , ..., e _n At this time, the matrix whose diagonal component is the square root of the eigenvalue is

And the matrix with eigenvectors in the column vector is E = (e ₁ ^T e ₂ ^T ... e _n ^T ), the whitening matrix V for the elements of the set A 'can be calculated with V = DE ^T it can. The whitening matrix calculation method is not limited to one, and a whitening matrix obtained by another calculation method can also be used in the present invention. For example, UV ′ obtained by multiplying a certain whitening matrix V ′ by an arbitrary orthogonal matrix U is also a whitening matrix. Since the matrix E is an orthogonal matrix, EDE ^T is also a whitening matrix. Therefore, EDE ^T may be used as the whitening matrix V instead of DE ^T.

  After the origin is translated, the normal area is determined as a spherical area with a radius R or less centered on the origin in the mapping space linearly transformed from the feature space by performing conversion using the whitening matrix V. (FIG. 4 (c)).

Next, the inspection object is inspected according to the following procedure.
The switching unit 31 in the preprocessing unit 30 is operated, and the switching unit 31 in the preprocessing unit 30 is operated so that the feature vector data obtained by the inspection unit 20 is sent to the determination unit. Then, the same processing as the processing shown for the normal product is performed on the inspection object, and its feature vector is generated (steps S12 to S18). Let Y be the feature vector of the inspected object obtained by these steps.

If the feature vector Y of the inspection object is included in the normal product region determined in step S11, it is determined as a normal product, and if not included, it is determined as a defective product (step S19). Here, the determination in step S19 can be performed by the following equation using the whitening matrix V used in step S11.
‖ V (Y−X _a ) ≦ ≤ R
That is, if the above equation holds, the feature vector Y of the object to be inspected is included in the normal product region, and is thus determined as a normal product.

  In addition to the above-described whitening method, the region determination unit 33 determines a normal product region using a method in the field of machine learning or data mining (support vector machine, naive Bayes classifier, cluster analysis, etc.), etc. A method that can properly capture the distribution of normal products can be used.

DESCRIPTION OF SYMBOLS 10 ... Measurement part 11 ... X-ray irradiation part 12 ... X-ray detection part 20 ... Inspection part 21 ... Image generation part 22 ... Image processing part 23 ... Feature-value extraction part 24 ... Feature vector generation part 30 ... Pre-processing part 31 ... Switching Unit 32 ... Learning unit 33 ... Area determination unit 40 ... Determination unit

Claims (8)

A step of generating an X-ray image of the article by irradiating the article with X-rays and measuring the amount of transmission, and extracting n types (n is an integer of 2 or more) of feature quantities from the X-ray image And a step of generating an n-dimensional feature vector based on the extracted n types of feature quantities,
generating feature vectors for m (m is an integer of 2 or more) normal products, and storing the distribution of feature vectors of the normal products as learning data;
Determining a normal product region based on the learning data;
Generating a feature vector for the inspection object, and determining whether the inspection object is a normal product according to whether the feature vector of the inspection object is included in the normal product region;
An X-ray inspection method characterized by comprising:   The X-ray inspection method according to claim 1, wherein the normal product region is provided on a mapping space obtained by converting a feature space having each feature parameter as a coordinate axis with a whitening matrix. The feature vectors of the m normal products are X ₁ , X ₂ , ..., X _m , the average of these feature vectors is X _a , the vector X _j ′ is X _j ′ = X _j −X _a , and the set When A ′ is A ′ = {X ₁ ′, X ₂ ′, ..., X _m ′}, the whitening matrix is the eigenvalue d ₁ , d _{2 of the} covariance matrix C for the elements of the set A ′. , ..., d _n and eigenvectors e ₁ , e ₂ , ..., e _n

The normal product region for the feature vector Y of the object to be inspected is
‖V (YX _a ) ‖ ≦ R
The X-ray inspection method according to claim 2, wherein R is a radius of the normal product region.   The X-ray inspection method according to claim 1, wherein the number m of the normal products is larger than the number n of types of the feature values. Image generation means for generating an X-ray image of the article by irradiating the article with X-rays and measuring the amount of transmission, and n types (n is an integer of 2 or more) of feature quantities from the X-ray image In an X-ray inspection apparatus comprising: a feature quantity extracting unit that extracts a feature vector; and a feature vector generating unit that generates an n-dimensional feature vector based on the n types of extracted feature quantities.
learning means for generating feature vectors for m (m is an integer of 2 or more) normal products, and storing the distribution of the feature vectors of the normal products as learning data;
Based on the learning data, a region determining means for determining a normal product region;
Determining means for generating a feature vector for the object to be inspected, and determining whether the object to be inspected is a normal product according to whether the feature vector of the object to be inspected is included in the normal product region; ,
An X-ray inspection apparatus characterized by comprising:   The X-ray inspection apparatus according to claim 5, wherein the normal product region is provided on a mapping space obtained by converting a feature space having each feature parameter as a coordinate axis with a whitening matrix. The feature vectors of the m normal products are X ₁ , X ₂ , ..., X _m , the average of these feature vectors is X _a , the vector X _j ′ is X _j ′ = X _j −X _a , and the set When A ′ is A ′ = {X ₁ ′, X ₂ ′, ..., X _m ′}, the whitening matrix is the eigenvalue d ₁ , d _{2 of the} covariance matrix C for the elements of the set A ′. , ..., d _n and eigenvectors e ₁ , e ₂ , ..., e _n

The normal product region for the feature vector Y of the object to be inspected is
‖V (YX _a ) ‖ ≦ R
The X-ray inspection apparatus according to claim 6, wherein R is a radius of the normal product region.   The X-ray inspection apparatus according to claim 5, wherein the number m of the normal products is larger than the number n of types of the feature values.

Images