JP2011014152A - Abnormal region detecting device and abnormal region detection method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a high speed and general abnormal region detecting device for highly precisely detecting the presence and the position of abnormality by using higher-order local autocorrelation features.SOLUTION: The abnormal region detecting device includes: a means for extracting feature data by higher-order local autocorrelation by pixel from image data; a means for adding the feature data about a pixel group by pixel in a prescribed range including each pixel; a means for calculating an index showing the abnormalities of the feature data for a partial space showing a normal region; a means for determining abnormality on the basis of the index; and a means for outputting a pixel position determined to be abnormal. A plurality of pieces of higher-order local autocorrelation feature data that differ due to displacement width may be extracted. Furthermore, the abnormal region detection device may be provided with a means for searching a partial space showing a normal region, based on main component vectors by a main component analyzing method from the feature data. Thus, determination of abnormalities according to pixels can be made, and accurate detection the position of abnormal regions can be made.

Description

  The present invention relates to an abnormal region detection apparatus and an abnormal region detection method for capturing an image and automatically detecting a region different from the normal region.

Conventionally, for example, in a defect inspection of a film of a flexible wiring board, when a wire is disconnected, it can be detected by energization, but thinning of a wire that may cause a product defect cannot be detected by energization. Therefore, it is necessary to detect abnormalities such as thinning visually or using an image.
In the case of visual observation, since the line is fine, it is necessary to use a visual inspection device or the like for enlarging the image. In addition, in order to provide feedback to the defect manufacturing process, it is necessary to output the coordinates and degree of the defect, which is troublesome and it is difficult to inspect all of a large number of products.

Therefore, today, an abnormality inspection using an image is automated in many product inspections. As an inspection method, for example, a pattern matching method is used in which a reference image registered for each product is compared.
On the other hand, various techniques have been proposed in order to detect a specific figure or the like from image data and to determine matching / mismatching with a registered image. The following patent document 1 filed by the present inventors discloses a technique of a learning adaptive image recognition / measurement method using a high-order local autocorrelation feature (hereinafter also referred to as HLAC data) for a two-dimensional image. Yes.

Japanese Patent No. 2982814

The pattern matching method, which is the conventional abnormal region detection method described above, has a problem that there is no versatility or learning effect on the target, and the position and direction need to be matched, the processing time is long, and the accuracy is poor. Moreover, since the conductor part of the flexible wiring board is a resin printed on a film and is granular, there is a problem that it is difficult to detect a figure such as a straight edge from an image and it is difficult to detect a defect.
On the other hand, in the learning adaptive image recognition method using higher-order local autocorrelation features, there is position invariance that results in the same detection result regardless of the position of the target (defect region) in the image. ) Cannot be specified, so that it cannot be applied to defect inspection.

  The object of the present invention is to solve the problems of the conventional example as described above, and to use a high-order local autocorrelation feature to detect the presence and absence of an abnormality and a position with high accuracy and a general-purpose abnormal area detection device and an abnormality It is in providing a region detection method.

  An abnormal region detection apparatus according to the present invention includes a feature data extraction unit that extracts feature data for each pixel from image data by high-order local autocorrelation, and a predetermined range including the pixel for each pixel separated by a predetermined distance. The pixel correspondence feature data generation means for adding the feature data extracted by the feature data extraction means for a pixel group, and the abnormality of the feature data generated by the pixel correspondence feature data generation means for a partial space indicating a normal region An index calculation unit that calculates an index to be indicated; an abnormality determination unit that determines that an abnormality occurs when the index is greater than a predetermined value; and an output that outputs a determination result that abnormally determines a pixel position that the abnormality determination unit determines to be abnormal The main feature is that it is provided with means.

  In the abnormal region detection apparatus described above, the feature data extraction unit is characterized in that it extracts a plurality of higher-order local autocorrelation feature data having different displacement widths. Further, the above-described abnormal area detecting device is characterized in that the index indicating the abnormality with respect to the partial space includes either information on the distance or angle of the characteristic data with respect to the partial space.

  Further, the abnormal region detection apparatus described above further includes a principal component subspace generation unit for obtaining a subspace indicating a normal region based on a principal component vector from the feature data extracted by the feature data extraction unit by a principal component analysis method. There is also a feature in the point. The above-described abnormal region detection apparatus is characterized in that the principal component subspace generation means obtains a subspace based on a principal component vector by a sequential principal component analysis method.

  Further, in the above-described abnormal area detection device, further, an index of similarity based on a canonical angle between the partial space obtained from the pixel correspondence feature data generated by the pixel correspondence feature data generation unit and the partial space, Classifying means for classifying each pixel using a clustering method, the principal component subspace generating means adds the feature data for each class to obtain a class-corresponding subspace, and the index calculating means includes: There is also a feature in that an index indicating abnormality of the feature data generated by the pixel-corresponding feature data generation unit with respect to the class-corresponding partial space is calculated.

  In the abnormal region detection method of the present invention, the feature data is extracted from the image data by high-order local autocorrelation for each pixel, and for each pixel separated by a predetermined distance, the pixel group in a predetermined range including the pixel is described above. A step of adding feature data; a step of calculating an index indicating abnormality of the feature data with respect to a partial space indicating a normal region; a step of determining abnormality when the index is greater than a predetermined value; a pixel determined to be abnormal The main feature is that it includes a step of outputting a determination result indicating that the position is abnormal.

The present invention has the following effects.
(1) Abnormality determination can be performed in correspondence with pixels, and the position of the abnormal region can be accurately detected.
(2) Conventionally, when there are a large number of objects, the abnormality detection accuracy decreases. However, if a predetermined range centered on pixels is appropriately selected, even if there are a large number of detection objects, the determination accuracy of the abnormal region is high. Will not drop.
(3) Since the amount of calculation for feature extraction and abnormality determination is small and the amount of calculation is constant regardless of the object, high-speed processing is possible.

(4) Since the normal region is statistically learned without being explicitly defined, it is not necessary to define what the normal region is at the design stage, and detection according to the monitoring target can be performed. Furthermore, the assumption about the monitoring target is unnecessary, and normality and abnormality can be discriminated for various monitoring targets, and the versatility is high. In addition, by updating the partial space of the normal area simultaneously with the abnormality determination, it is possible to follow the change of the normal area.
(5) Even if there are a plurality of types of objects, the detection accuracy is further improved by classifying them by position. Further, the classification may be processed in advance, or the classification process can be automatically updated simultaneously with the abnormality determination.

It is a block diagram which shows the structure of the abnormal area | region detection apparatus by this invention. It is explanatory drawing which shows the outline | summary of the abnormal area | region detection process by this invention. It is explanatory drawing which shows the autocorrelation process coordinate in a two-dimensional pixel space. It is explanatory drawing which shows the content of an autocorrelation mask pattern. It is a flowchart which shows the content of the abnormality detection process of this invention. It is a flowchart which shows the content of the pixel corresponding | compatible HLAC data generation process. It is a flowchart which shows the content of the HLAC characteristic data generation process. It is explanatory drawing which shows the property of the subspace of HLAC characteristic. It is explanatory drawing which shows the image showing an input image and an abnormality determination result.

  In this specification, an abnormal area is defined as “not a normal area”. If the normal region is a region where the distribution is concentrated when a statistical distribution of region features is considered, the normal region can be learned as a statistical distribution without a teacher. An area that greatly deviates from the distribution is defined as an abnormal area.

  As a specific method of abnormal region detection, a partial space of normal region features is generated in a region feature space based on higher-order local autocorrelation features, and the abnormal region is detected using the distance or angle from the partial space as an index. . For example, a principal component analysis method is used to generate the normal region subspace, and the principal component subspace is configured by a principal component vector having a cumulative contribution rate of 0.99, for example.

  Here, the high-order local autocorrelation feature has a property that it does not require extraction of an object and is additive in the screen. Due to this additivity, when a normal area subspace is configured, the feature vector will fit in the normal area subspace no matter how many normal wires are in the screen, and there is an abnormal area in one of them. Then, it jumps out of the partial space and can be detected as an abnormal value. Since it is not necessary to track and cut out the lines individually, the calculation amount is constant without being proportional to the number of lines, and can be calculated at high speed.

Further, in the present invention, in order to detect the position of the object, HLAC data of a predetermined area including (centering) this pixel for each pixel separated by a predetermined distance (arbitrary distance of one pixel or more) The pixel-corresponding HLAC feature data obtained by integrating is obtained, and abnormality determination is performed based on the distance or angle between this data and the normal region partial space. With this process, it is possible to determine normality / abnormality for each pixel.
Higher order local autocorrelation (HLAC) features are used to extract region features from image data. The k-th component of the HLAC feature is given by Equation 1 below.

Here, I (r) represents an image, and a variable r (reference point) and N local displacements a _N ^k are two-dimensional vectors having in-screen coordinates x and y as components. B _N ^k is a local displacement matrix having N local displacements as column components. Further, the integration range is a W × H image area, and W and H represent the width and height of the image.

  Further, in the present invention, the high-order correlation of a wider region is also extracted than the 3 × 3 neighborhood region for the HLAC feature. Therefore, using the matrix R, a solid higher-order local autocorrelation feature formulated as follows is considered.

This feature is obtained by reducing the scale in the image plane by 1 / λ _x times in the horizontal direction and 1 / λ _y times in the vertical direction with respect to the image I (r). I _amp (r, R) = I (Rr The feature amount is proportional to the feature amount extracted from (). That is, the CHLAC feature h _amp (Br, R) extracted for I _amp (r, R) is as follows.

Therefore, it can be seen that the feature quantity that has been correlated in a wider range using the matrix R is λ _x λ _y times the HLAC feature quantity extracted from the reduced image. In fact, if you want to extract more correlation features from a target, you need to change the scale of the image appropriately according to the target and the image. So that the scale can be adjusted.

Therefore, in order to be robust with respect to the scale, features (λ _x , λ _y ) with different scales are added to the vector components and used as new features. For example, by combining the feature extracted with (λ _x , λ _y ) = (1, 1) and the feature extracted with (λ _x , λ _y ) = (2, 2), it is robust to a 35 × 2 dimensional scale. Get features.

  FIG. 1 is a block diagram showing a configuration of an abnormal area detecting apparatus according to the present invention. For example, the digital camera 10 outputs image data of a product to be inspected. It may be a camera incorporated in a microscope. The digital camera 10 may be a monochrome camera or a color camera. The computer 11 may be a known personal computer (PC) having an input terminal for capturing an image such as a USB. The present invention is realized by creating, installing, and starting a processing program to be described later in any known computer 11 such as a personal computer.

  The monitor device 12 is a well-known output device of the computer 11 and is used, for example, to display to the operator that an abnormal region has been detected. The keyboard 13 and the mouse 14 are well-known input devices used for input by the operator. The digital camera 10 may be connected to the computer 11 via an arbitrary communication network, or data may be transferred to the computer 11 via a memory card.

  FIG. 2 is an explanatory diagram showing an outline of the abnormal area detection processing according to the present invention. For example, for 360 × 240 pixels, 256 gray scale input image data (a), first, displacement width 1, that is, the displacement width is set equal to the pixel width, and pixel-corresponding HLAC data is calculated for each pixel (b) ). The HLAC data will be described later. Next, the displacement width is set to 2, that is, twice the pixel width, and pixel-corresponding HLAC data is calculated for each pixel in the same manner (b). This process is repeated up to the maximum displacement width (for example, 3). As a result, pixel-corresponding HLAC data (c) for each displacement width is obtained.

  Next, a set of feature data obtained by adding (g) HLAC data (c) for each displacement width is defined as all HLAC feature data (h). Then, a principal component subspace is obtained from all HLAC feature data (h) by principal component analysis or sequential principal component analysis (i). Usually, since most areas of the image are normal areas, this principal component subspace represents the characteristics of the normal areas.

  On the other hand, from the pixel-corresponding HLAC data (c) for each displacement width, a process of adding HLAC data of a predetermined area (for example, 10 × 10) centered on the target pixel while moving the target pixel for each displacement width is performed. (E) Pixel corresponding HLAC feature data (f) is obtained. Finally, abnormality determination is performed for each pixel based on the distance or angle between the normal subspace and the pixel-corresponding HLAC feature data (j), and the pixel position determined to be abnormal is displayed and output as an abnormal area (k).

  In the present invention, the normal region partial space generation processes (g) and (i) are performed on the entire region of the image in advance or on a part of the region sampled randomly or regularly. The abnormality determination process (j) may be performed based on the partial space information.

  Details of the processing will be described below. FIG. 5 is a flowchart showing the contents of the abnormality detection process of the present invention. For example, assume that 256 gray scale image data has already been read. In S10, the displacement width λ is set to 1. In S11, pixel-corresponding HLAC data based on the displacement width λ is generated and stored for each pixel of the image. Details of this processing will be described later.

  FIG. 3 is an explanatory diagram showing autocorrelation processing coordinates in a two-dimensional pixel space. In the present invention, correlation is performed for pixels in a square of 3 × 3 = 9 pixels centering on the reference pixel of interest (when λ = 1). The mask pattern is information indicating a combination of pixels to be correlated, and the pixel of interest (reference point) that is the center of the square is always selected, and the surroundings represent local displacement. The pixel data selected by the mask pattern is used to calculate the correlation value.

  FIG. 4 is an explanatory diagram showing the contents of the autocorrelation mask pattern. FIG. 4 (1) shows the simplest zeroth-order mask pattern (one) with only the target pixel. (2) is an example of a primary mask pattern in which two pixels are selected (a total of five), and the number in the frame indicates the number of times that the pixel value is multiplied. (3) The subsequent rows are examples of a tertiary mask pattern (29 in total) in which three pixels are selected. Excluding overlapping patterns when the pixel of interest is moved, the total number of mask patterns for grayscale images, that is, HLAC feature components, is 35. That is, the high-order local autocorrelation feature vector for one two-dimensional data is 35 dimensions.

  In S12, the HLAC feature data corresponding to the λ-corresponding pixel is generated by adding the HLAC data of a predetermined area centered on the target pixel, for example, a 10 × 10 area, while moving the target pixel. Details of this processing will be described later. In S13, all the pixel-corresponding HLAC data are added, and the λ-corresponding all HLAC feature data is stored.

  In S14, 1 is added to λ. In S15, it is determined whether or not λ has exceeded the maximum value (for example, 3). If the determination result is negative, the process proceeds to SS11, but if the determination result is affirmative, the process proceeds to SS16. In S16, for each of the λ corresponding HLAC feature data and the λ corresponding all HLAC feature data, all λ are collected. Accordingly, when the maximum value of λ is 3, the dimensions of the pixel-corresponding HLAC feature data and the total HLAC feature data are 35 × 3 = 105 dimensions.

  In S17, a principal component vector is obtained from all HLAC feature data by principal component analysis or sequential principal component analysis method, and is set as a subspace of a normal region. Since the principal component analysis method itself is well known, an outline will be described. First, in order to construct a subspace of the normal region, a principal component vector is obtained by principal component analysis from all HLAC feature data. The M-dimensional HLAC feature vector x is expressed as follows.

  Note that M = 35. A matrix U (eigenvector) in which principal component vectors are arranged in columns is expressed as follows.

The matrix U in which the principal component vectors are arranged in columns is obtained as follows. The autocorrelation matrix R _x is shown in the following equation.

The matrix U is obtained from the eigenvalue problem of the following equation using this autocorrelation matrix _Rx .

  The eigenvalue matrix Λ is expressed by the following equation.

The cumulative contribution rate α _K up to the Kth eigenvalue is expressed as follows.

Therefore, the space spanned by the eigenvectors u ₁ ,..., U _K up to a dimension where the cumulative contribution rate α _K becomes a predetermined value (for example, α _K = 0.99) is applied as a subspace of the normal region. Note that the optimum value of the cumulative contribution rate α _K depends on the monitoring target and the detection accuracy, and is therefore determined by experiments or the like. By performing the above calculation, a partial space corresponding to the normal region is generated.

  Next, a sequential principal component analysis method for obtaining subspaces incrementally without solving the eigenvalue problem and obtaining the covariance matrix will be described. Since real-world applications handle large amounts of data, it is difficult to keep all the data. Therefore, the subspace of the normal region is learned and updated sequentially.

As a method considered as a sequential principal component analysis, first, an eigenvalue problem may be solved for each step. The autocorrelation matrix R _x required for the eigenvalue problem is updated as follows.

Here, R _x (n) is an autocorrelation matrix at the nth step, and x (n) is an input vector at the nth step. Although this is faithful to the principal component analysis method described above, there is a problem that the calculation amount is large because the eigenvalue problem must be solved for each step. Therefore, CCIPCA, which is a technique for sequentially updating eigenvectors without solving the eigenvalue problem without obtaining a correlation matrix, is applied. The contents of CCIPCA are disclosed in Non-Patent Document 1 below.
Juyang Weng, Yilu Zhang and Wey-Shiuan Hwang, "Candid Covariance-Free Incremental Principal Component Analysis", IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.25, No.8, pp.1034-1040, 2003.

  This algorithm is a very fast method because it is not necessary to solve the eigenvalue problem for each step. In this method, eigenvalue convergence is not so good, but eigenvector convergence is fast. The first eigenvector and the first eigenvalue are updated as follows.

Here, the eigenvector is v / ‖v‖, and the eigenvalue is ‖v‖. This updating rule proves that v (n) → ± λ ₁ e ₁ when n is infinite. Here, λ ₁ is the maximum eigenvalue of the correlation matrix R of the sample, and e ₁ is a corresponding eigenvector. It is shown that the n-th eigenvector and the n-th eigenvalue are updated incrementally from the first eigenvector and the first eigenvalue according to the Gram-Schmidt orthogonalization method and converge to the true eigenvalue and eigenvector. The detailed update algorithm is as follows.

  In the present invention, the upper limit value is determined instead of obtaining M in all dimensions for the eigenvector obtained by CCIPCA. When solving the eigenvalue problem, the eigenvalue is calculated and then the cumulative contribution rate is obtained, and the cumulative contribution rate is taken to a dimension exceeding, for example, 0.99999. First, the conventional method requires a large amount of calculation. In order to obtain the contribution rate, all eigenvalues must be estimated, and the calculation time takes several tens of seconds even if the calculation time for feature extraction is not included. On the other hand, if the dimension is calculated as a constant, for example, 4, it can be calculated in several milliseconds.

  The second reason is that the eigenvalue convergence is slow in this CCIPCA method. When the CCIPCA method is used for the number of data of several thousand frames, it can be seen that the dimension of the normal region subspace is finally about two hundred and does not converge to the convergence value of 4 at all. For these reasons, the dimension of the subspace is determined to be constant. The approximate value of this parameter can be obtained by solving the eigenvalue problem once for an input vector of a certain time width.

  In S18, the distance d⊥ between the pixel corresponding HLAC feature data obtained in S16 and the partial space obtained in S17 is obtained.

  FIG. 8 is an explanatory diagram showing the properties of the subspace of the HLAC feature. In FIG. 8, for the sake of simplicity, the HLAC feature data space is two-dimensional (actually 26 × 3 dimensions), and the normal area subspace is one-dimensional (in the embodiment, for example, the cumulative contribution rate is 0.99, 3). The HLAC feature data of the normal area forms a group for each number of monitoring targets.

  The normal area subspace S obtained by the principal component analysis exists in the vicinity including the HLAC feature data of the normal area. On the other hand, the HLAC feature data A of the abnormal area has a large vertical distance d⊥ with respect to the normal area partial space S. Accordingly, the abnormal region can be easily detected by measuring the vertical distance d⊥ between the HLAC feature data and the partial space of the normal region.

The distance d⊥ is obtained as follows. The projector P onto the normal subspace spanned by the principal component orthogonal basis U _K = [u ₁ ,..., U _K ] and the projector P⊥ onto the orthogonal complement space are as follows.

U ′ is a transposed matrix of the matrix U, and I _M is an M-th unit matrix. The square distance in the orthogonal complement space, that is, the square distance d ²の of the perpendicular to the subspace U can be expressed as follows.

  In the present embodiment, this vertical distance d⊥ can be used as an indicator of whether or not it is abnormal. However, the above-mentioned vertical distance d 指標 is an index that varies depending on the scale (norm of the feature vector). Therefore, there is a risk that different determination results may be obtained depending on the difference in scale. Therefore, as another index, the following more robust index may be adopted.

  First, consider the case where the angle with the subspace S, that is, sin θ is used as an index. However, this index is an index that gives a very large value even for features with a very small scale such as noise, and is not a very appropriate index. Therefore, in order to obtain a small value even when the scale is small, this index is changed as follows.

  Here, c is a positive constant. With this index, the abnormality determination value is corrected with respect to the scale and becomes an index that is also resistant to noise. This index means that an angle from a point shifted in the horizontal axis direction by −c from the origin on the graph of FIG. 8 is measured.

  In S19, it is determined whether or not the distance d⊥ exceeds a predetermined threshold value. If the determination result is negative, the process proceeds to S20, but if the determination result is affirmative, the process proceeds to S21. In S20, it is determined as normal. Moreover, in S21, it determines with it being abnormal. In S22, it is determined whether or not the determination process has been completed for all the pixels. If the determination result is negative, the process proceeds to S18, but if the determination is affirmative, the process proceeds to S23. In S23, the determination result is output.

  FIG. 6 is a flowchart showing the contents of the pixel-corresponding HLAC data generation process of S11. In S30, the correlation pattern corresponding feature value is cleared. In S31, one unprocessed pixel (reference point) is selected. In S32, one unprocessed pattern is selected. In S33, based on the correlation pattern and the displacement width λ, the correlation value is calculated by multiplying the pixel luminance value at the position corresponding to the pattern by using Equation 1 described above. This process corresponds to the calculation of f (r) f (r + a1)... F (r + aN) in the above-described equation 1.

  In S34, the correlation value is stored in correspondence with the correlation pattern. In S35, it is determined whether or not processing has been completed for all patterns. If the determination result is negative, the process proceeds to S32, but if the determination is affirmative, the process proceeds to S36. In S36, the correlation value group is stored for each pixel. In S37, it is determined whether or not the processing has been completed for all pixels. If the determination result is negative, the process proceeds to S31, but if the determination is affirmative, the process proceeds to S38. In S38, a set of correlation value groups is output as pixel-corresponding HLAC data.

  FIG. 7 is a flowchart showing the contents of the HLAC feature data generation process of S12. In S40, one unprocessed pixel (reference point) is selected. Although all pixels may be scanned as a selection method, selection (sampling) may be performed for each pixel separated by a predetermined distance of 2 pixels or more on the XY coordinates of the image. In this way, the processing amount is reduced.

  In S41, pixel-corresponding HLAC data in a predetermined area centered on the reference point is added. The predetermined area may be, for example, 10 × 10 including the pixel of interest (centered). This processing corresponds to the integration operation in Equation 1 described above, in which the pixel of interest is moved (scanned) by a desired range and the correlation values are added (correlation values are added for each dimension).

  In S42, the added data is stored in correspondence with the pixels. In S43, it is determined whether or not processing has been completed for all pixels. If the determination result is negative, the process proceeds to S40, but if the determination is affirmative, the process proceeds to S44. In S44, the set of feature addition values is output as pixel-corresponding HLAC feature data.

  FIG. 9 is an explanatory diagram illustrating an input image and an image representing an abnormality determination result. FIG. 9A shows the input grayscale image, and FIG. 9B shows the grayscale obtained by normalizing the abnormal index value for each pixel processed by the method of the present invention with the maximum value being 255 and the minimum value being 0. Represents an image. The index value of the central defect portion is large (white), indicating that a defect has been detected.

  In the first embodiment, the partial space of the normal region is obtained from the entire image. However, in the second embodiment, for example, a region having a straight wiring and a region having many round patterns are mixed. In this case, the region is classified into a region having a straight line and a region having a large number of round patterns, and abnormality determination is performed by obtaining a partial space of a normal region for each class. By doing so, the determination accuracy is improved.

  First, a principal component vector is obtained from all HLAC data and pixel-corresponding HLAC feature data by a principal component analysis method or a sequential principal component analysis method. Next, the canonical angles of the two obtained principal component vectors are obtained, and the pixels are classified into classes according to the similarity based on the canonical angles.

The canonical angle means an angle formed by two subspaces in statistics, and N (≦ M) canonical angles can be defined between the M-dimensional subspace and the N-dimensional subspace. The second canonical angle θ ₂ is the minimum angle measured in the direction orthogonal to the minimum canonical angle θ ₁ , and similarly the third canonical angle θ ₃ is the minimum angle measured in the direction orthogonal to θ ₂ and θ _1. . The basis vectors of the subspaces L ₁ and L ₂ in the F-dimensional feature space are Φ _i and ψ _i , and the F × F projection matrix calculated from these is shown below.

The i-th largest eigenvalue λ _{i of} P ₁ P ₂ or P ₂ P ₁ is cos ² θ _i . The relationship between the M-dimensional subspace L ₁ and the N-dimensional subspace L ₂ is completely defined by N canonical angles. If the two subspaces are completely coincident, all N canonical angles are 0 degrees. As the distance between them increases, the lower canonical angle increases, and when they are completely orthogonal, all angles are 90 degrees. Thus, a plurality of canonical angles represents the structural similarity between the two subspaces. Therefore, the similarity S [n] is defined as follows using n (≦ N) canonical angles, and this index is used.

Next, clustering is performed using the Mean Shift method on the index value obtained by the similarity of the canonical angle. The contents of the Mean Shift method are disclosed in Non-Patent Document 2 below. The Mean Shift method is a clustering method that does not give the number of classes, and it is necessary to set a scale parameter indicating how close the neighborhood is. In this embodiment, since the similarity of the canonical angle having a value between 0 and 1 is used as an index, the scale parameter is set to about 0.1.
Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 24, No. 5, pp. 603-619, 2002.

  Finally, pixel-corresponding HLAC data is added for each class, the principal component vector for each class is obtained from the added HLAC feature data by the principal component analysis method or the sequential principal component analysis method as described above, and the normal region for each class The subspace of Then, the abnormality determination is performed based on the distance between the pixel-corresponding HLAC feature data and the partial space of the normal region corresponding to the determined class.

  Although the embodiment for detecting an abnormal region has been described above, the following modifications may be considered in the present invention. In the embodiment, an example in which an abnormal region is detected while updating a partial space of a normal region has been disclosed. However, a normal region partial space is generated in advance by a learning phase, and an abnormality is detected using a fixed partial space. An area may be detected. Furthermore, a small amount of data may be learned in advance by random sampling or the like.

  In the embodiment, an example in which feature data is generated for each pixel has been disclosed. However, feature data closer in position are more similar. Therefore, for example, by performing the process shown in FIG. 9 for each pixel separated by a predetermined distance, the processing load is reduced and the area can be made faster. However, this method is a trade-off with the problem of specifying the location or the detection accuracy, and an appropriate setting is required for each problem.

  In the embodiment, an example in which a value that is an integer multiple of a pixel is used as λ, which is a parameter for controlling the scale, is disclosed, but any real value including a decimal point can be used as λ. However, pixel values need to be interpolated in the case of real values. Further, the feature may be extracted after the image is enlarged / reduced at an arbitrary magnification including a decimal point.

DESCRIPTION OF SYMBOLS 10 ... Digital camera 11 ... Computer 12 ... Monitor apparatus 13 ... Keyboard 14 ... Mouse

Claims (6)

For each pixel from image data, a high number of sets of data corresponding to the number of correlation patterns is generated by multiplying pixel luminance values at a plurality of positions corresponding to the pattern based on each of a plurality of predetermined correlation patterns. Feature data extraction means for extracting feature data by next local autocorrelation;
For each pixel separated by a predetermined distance in the image data, pixel-corresponding feature data is generated by adding the feature data extracted by the feature data extraction unit for a pixel group in a predetermined range including the pixel. Feature data generation means;
A normal subspace generating means for obtaining a subspace indicating a normal area based on a principal component vector from a feature data extracted by the feature data extracting means by a principal component analysis method;
This is an index indicating the abnormality of the pixel-corresponding feature data with respect to the partial space indicating the normal region, and includes either distance or angle information between the pixel-corresponding feature data and the partial space indicating the normal region. An index calculation means for calculating an index;
Abnormality determining means for determining an abnormality when the index is greater than a predetermined value;
An abnormality area detection apparatus comprising: output means for outputting a determination result that makes the pixel position determined abnormal by the abnormality determination means abnormal.   The abnormal region detection apparatus according to claim 1, wherein the feature data extraction unit extracts a plurality of higher-order local autocorrelation feature data having different displacement widths. The abnormal area detection apparatus according to claim 1, wherein the image data is a grayscale grayscale image.   2. The abnormal region detection apparatus according to claim 1, wherein the normal subspace generation unit obtains a subspace based on a principal component vector by a sequential principal component analysis method in which an upper limit value is provided for a dimension. Further, an index of similarity based on a canonical angle between the partial space obtained from the pixel-corresponding feature data generated by the pixel-corresponding feature data generating unit and the partial space indicating the normal region is obtained, and a pixel is obtained using a clustering method. Classifying means for classifying every time,
Class subspace generation means for adding the feature data extracted by the feature data extraction means for each class, and obtaining a subspace for each class indicating a normal area based on a principal component vector by a principal component analysis method,
The index calculation means includes feature data as an index indicating abnormality of feature data generated by the pixel-corresponding feature data generation means with respect to a class-corresponding partial space indicating the normal area, The abnormal region detection device according to claim 1, wherein either the distance information or the angle information is calculated. For each pixel from the image data, the computer multiplies pixel luminance values at a plurality of positions corresponding to the pattern based on each of a plurality of predetermined correlation patterns to obtain a set of data corresponding to the number of correlation patterns. Extracting feature data by higher order local autocorrelation to generate,
A computer generates pixel-corresponding feature data by adding the feature data extracted by the feature data extraction unit for a pixel group in a predetermined range including the pixel for each pixel separated by a predetermined distance in the image data. Step to do,
A step of obtaining a partial space indicating a normal region based on a principal component vector by a principal component analysis method from all feature data obtained by adding the feature data extracted by the feature data extraction means;
The computer is an index indicating abnormality of the pixel-corresponding feature data with respect to the partial space indicating the normal region, and information on either distance or angle between the pixel-corresponding feature data and the partial space indicating the normal region Calculating the metrics it contains,
A step of determining that the computer is abnormal when the index is larger than a predetermined value;
A method for detecting an abnormal region, comprising: a step of outputting a determination result that the computer determines that the pixel position determined to be abnormal is abnormal.

Images