JP6099479B2 - Crack detection method - Google Patents

Description

  The present invention relates to a crack detection method for detecting cracks generated on a concrete surface, and in particular, even when it is difficult to detect cracks due to dirt or lighting conditions of the photographed concrete surface, it is simple and highly accurate. The present invention relates to a crack detection method capable of detecting cracks.

  As a method for detecting cracks on the concrete surface, conventionally, a method in which an investigator visually observes using a scale and measures the width and length of the crack has been common. However, this visual observation method requires a large amount of labor and time to accurately process a large amount of information when there are large variations in accuracy due to the measurement skill of the investigator and when there are a large number of cracks. There was a problem.

  In order to solve the above problem, an image processing technique is applied in which a captured image of a concrete surface is taken into a computer, and the image is binarized into a cracked area and other areas. The image binarization process is to express the image density to 0 or 1 with respect to a certain density value. For example, 2 obtained by the binarization process for the input image f (i, j). The binarized image b (i, j) is b (i, j) = 1 (f (i, j)> k) and 0 (f (i, j) ≦ k). Here, k is a threshold value for binarization, and therefore it can be said that the quality of the binarized image is determined by the selection of the threshold value k.

  As a conventional method for obtaining a threshold value, there is a processing method using a fixed threshold value or a variable threshold value. Examples of the processing method using a fixed threshold include a P tile method, a mode method, and a method using a correlation ratio. The processing method using a fixed threshold creates a density histogram of the target image, and a bimodal histogram is obtained in which a clear valley appears between the density value of the background (concrete surface) of the image and the density value of the crack. This is an effective method in some cases.

  On the other hand, the processing method using a variable threshold is an effective method in the case where shooting unevenness occurs due to illumination conditions or the like, and the density value of the background and the density value of the target portion are not constant for the entire image. This variable threshold processing method is a method in which an average density value of a local region centered on a pixel of interest is used as a threshold. A drawback of this method is that, for example, an image having a lot of noise other than cracks is generated in accordance with a subtle change in shading of the background area.

  In the conventional image processing method, a threshold is determined for a photographed input image, and cracks are extracted while performing a binarization process. That is, the general processing flow is as follows. 1) Capture the captured image into a computer and create an input image. 2) Correct the density of the input image. 3) Perform binarization and extract cracks.

  The conventional image processing method described above can detect a crack with relatively high accuracy in detecting a crack on a concrete surface having a uniform concentration. However, the surface of an actual concrete structure contains various stains, and the density of cracks generally varies depending on the width and depth of the cracks. When a conventional image processing method is used for such a concrete surface, various problems may occur when extracting cracks. For example, in the case of the fixed threshold processing, if the dirt region and the crack region on the concrete surface have similar density values, it is extremely difficult to binarize them. Even if the density histogram exhibits bimodality and the threshold value can be easily determined, there is a very high possibility that the area that is determined to be a cracked area includes a dirty area. On the other hand, if a new threshold value is set so as not to include the dirt region around the crack, another crack region will be excluded this time.

  In the case of variable threshold processing, as the dirt on the concrete surface increases, more noise other than cracks will be included in the cracked extracted image. It is impossible to determine at all whether or not the region is a cracked region.

  In order to solve the problems of the conventional methods described above, the present inventors can easily detect cracks with high accuracy even when cracks are difficult to detect due to dirt on the concrete surface or lighting conditions. A detection method is conceived and disclosed in Patent Document 1. This crack detection method calculates wavelet coefficients corresponding to two contrasted concentrations, calculates wavelet coefficients when the two concentrations are changed, creates a wavelet coefficient table, and is a crack detection target. A step of creating a wavelet image by inputting a captured image of the concrete surface into a computer and converting the input image into a wavelet. The average density of neighboring pixels in the local area and the density of the target pixel in the wavelet coefficient table. If the wavelet coefficient of the pixel of interest is greater than the threshold, the pixel of interest is determined to be cracked, and if it is less than the threshold, the pixel of interest is determined not to be cracked. Note while changing Creating a crack extracted image by performing a comparison with the wavelet coefficients of the pixel and the threshold value, the detection method comprising.

  In this detection method, a pixel that is connected to an arbitrary pixel is connected to an arbitrary pixel by an outline tracking process, and a set of connected pixels is labeled. . Here, when the number of constituent pixels (or corresponding area) for each labeling is smaller than the threshold value, the labeled aggregate is regarded as noise and removed.

  Although the detection accuracy is greatly improved by the crack detection method to which the above-described wavelet transform is applied, according to further verification on the crack detection method by the present inventors, the wavelet coefficient value and the wavelet coefficient table are used as analysis parameters. The number of pixels determined to be cracks obtained from the contour tracking process as a threshold for noise removal is 25, and the number of pixels determined as cracks is 10 (if this value is less than this value, the pixels of this block are removed). As a result of calculation, a lot of noise other than cracks has been detected, and it is known that further improvement is necessary.

  Although it is possible to reduce the noise part other than cracks by increasing the crack detection threshold or increasing the noise removal threshold, it is not possible to detect some fine cracks or discontinuous cracks. It is difficult to say that it is a good idea.

  Therefore, in order to solve the above-mentioned problems, the present inventors have used a method for multivariate analysis of areas such as cracks, formwork traces, and dirt on the concrete surface from a binarized image obtained by computer processing. It has come to the idea of automatically discriminating cracks and noise with high accuracy by a certain discrimination process.

  More specifically, as the feature amount related to the shape of each object (one block formed by connecting pixels having the same value in the binarized image) of the obtained binarized image, the area, the perimeter, the circle Select items such as degrees and calculate feature values. Regarding the feature quantities of these blocks, a feature quantity table is created in advance.

  The feature table may be created on a concrete surface that is currently subject to crack detection on the concrete surface, or on a concrete surface other than that. A binarized image is created, and a plurality of feature quantities of each block constituting the binarized image are calculated by a computer and tabulated. The feature table further includes observation results regarding this arbitrary concrete surface, that is, the classification “3” when each lump is cracked, and the classification “2” when the noise type is a formwork or a scar. ”In the case of noise such as stains and dirt, among the noise types, the classification“ 1 ”, etc., and so on, whether each lump as an observation result belongs to a group of cracks or noise types Assigned. Discriminant analysis is performed on this feature quantity table with a plurality of feature quantities of each block as explanatory variables, and cracks or one of the noise types as an objective variable. It is determined that it is one of the noise types, and the determined result is used as a predicted value.

  According to the improved crack detection method, the present inventors have specified that a high result of a discrimination rate between cracks and fine noises of 92% or more can be obtained.

  However, the discrimination rate between cracks and form marks is as low as about 70%, and it is also specified by the present inventors that there is a possibility of erroneous discrimination.

  The present inventors analyze this cause as follows. In other words, the Gabor function is applied to the crack detection method using the wavelet transform described above. This Gabor function is suitable for detecting a narrow linear object such as a crack, but has a wide width. This is because it is not suitable for detecting objects such as planar formwork traces, peeling parts, and water leakage parts. Thus, the discrimination rate can be increased by applying a Laplacian Gaussian function to the detection of objects such as wide planar formwork traces, peeled portions, and water leakage portions.

  Therefore, the present inventors have come up with the idea of a crack detection method that can accurately remove both surface noise such as form marks and fine noise, and can detect cracks with high accuracy. .

JP 2006-162583 A

  The present invention has been made in view of the above-described problems, and can remove both surface noise such as formwork traces and fine noise with high accuracy, and can detect cracks with high accuracy. The object is to provide a crack detection method.

In order to achieve the above object, a crack detection method according to the present invention comprises:
(Method A) A method of creating a crack discrimination image for detecting cracks occurring on the concrete surface, the method comprising Step 1A, Step 2A, and Step 3A,
(Method B) Plane objects represented by artificial structures such as stains, springs, peeling and peeling, and artificial structures such as cables and overhead lines, which are objects having a spread on the surface of the concrete. A non-crack discrimination image corresponding to the method, comprising: Step 1B, Step 2B, and Step 3B;
A crack detection method comprising (Method A) and (Method B),
(Method A)
Simulate the density of the crack and the concrete surface, apply the Gabor function and calculate the wavelet coefficients corresponding to the two concentrations to be compared, and the wavelet coefficients when these two concentrations are changed respectively. Step 1A for creating a wavelet coefficient table by calculating a wavelet coefficient table, inputting a captured image of a concrete surface to be cracked into a computer as an input image, and wavelet transforming the input image;
In the wavelet coefficient table, the average density of neighboring pixels in the local area assumed to be the density of the concrete surface and the wavelet coefficient corresponding to the density of the target pixel assumed to be the crack density are used as threshold values, and in any neighboring pixels When the wavelet coefficient of an arbitrary pixel of interest is larger than a threshold value, the pixel of interest in the neighboring pixel is determined as a cracked pixel, and the wavelet coefficient of an arbitrary pixel of interest in an arbitrary neighboring pixel is smaller than the threshold value. determines that no cracking of the target pixel in the near neighbor pixel, compares the wavelet coefficients and the threshold value of the target pixel while varying the local region and the pixel of interest, not further such to put the contour tracing processing, connecting the cracking pixel a single lump formed, the threshold of mass the mass is set in advance If remote small, it is determined that noise, by if greater than the threshold value of the mass is determined that cracking, steps 2A to create a binarized image in which a part of the noise is being removed,
Identify the feature quantity of the lump, refer to the feature quantity table in which cracks and noise types other than cracks are defined in advance for each feature quantity of the lump, use multiple feature quantities as explanatory variables, Perform a discriminant analysis with the noise type as the objective variable to determine that the mass with the specified feature quantity is either a crack or a noise type,
Step 3A, which creates a crack discriminating image in which noise is further removed from the binarized image by executing the identification of the feature amount for each chunk and discriminating and analyzing each chunk in Step 2A. ,
The feature amount table is created by executing steps 1A and 2A on an arbitrary concrete surface to create a binarized image, and a plurality of feature amounts of the respective blocks constituting the binarized image. It is calculated by computer and assigns one kind of group consisting of cracks and noise species, which is the observation result for each lump,
(Method B)
When the density of the object and the density of the concrete surface are set in a pseudo manner, the wavelet coefficients corresponding to the two densities to be compared are calculated by applying the Laplacian Gaussian function, and the two densities are changed respectively. A wavelet coefficient table is calculated and a wavelet coefficient table is created, and the captured image of the concrete surface that is the detection target of these objects is input to a computer as an input image, and a wavelet image is created by wavelet transforming this input image Step 1B,
In the wavelet coefficient table, the average density of neighboring pixels in the local region assumed to be the density of the concrete surface, and the wavelet coefficient corresponding to the density of the target pixel assumed to be the density of the planar object are set as threshold values, and any arbitrary When the wavelet coefficient of an arbitrary pixel of interest in a neighboring pixel is larger than a threshold, the pixel of interest in the neighboring pixel is determined as a planar object, and the wavelet coefficient of an arbitrary pixel of interest in an arbitrary neighboring pixel is smaller than the threshold In this case, it is determined that the target pixel in the neighboring pixel is not a planar object, the wavelet coefficient of the target pixel is compared with the threshold value while changing the local region and the target pixel, and the pixel determined as the planar object It performs determination processing of the planar object and noise to further perform the contour tracing process, the planar target And one mass is formed by connecting the determined pixel, if smaller than the threshold of mass the mass is set in advance is determined to noise, if greater than the threshold mass and the planar object Step 2B for creating a binarized image in which a part of noise is removed by determination,
Identify the feature quantity of the lump, refer to the feature quantity table that prescribes for each feature quantity of the lump, deformations such as spots, spring water and peeling off, artificial structures such as cables and overhead lines, etc. , Using a plurality of feature quantities as explanatory variables, performing a discriminant analysis with these planar objects as objective variables, and determining that the mass in which the feature quantities are specified is a planar object,
From Step 3B, which creates a non-crack discrimination image in which noise is further removed from the binarized image by performing the identification of the feature amount for each chunk and the discrimination analysis of each chunk for all the chunks in Step 2B. Become
Next, a crack image is created based on the difference between the crack discrimination image created by Method A and the non-crack discrimination image created by Method B.

  A wavelet means a small wave, and a basic unit of a localized wave can be expressed by an expression using a wavelet function. By enlarging or reducing the wavelet function, it is possible to simultaneously analyze time information, spatial information, and frequency information. When this wavelet coefficient is applied to a concrete surface having cracks, the characteristic of this coefficient is that it depends on the concentration of the concrete surface, the concentration of cracks, and the crack width. For example, the value of the wavelet coefficient tends to increase as the crack width increases, and the value of the wavelet coefficient tends to increase as the density of cracks increases (closer to black).

  An algorithm for detecting cracks using wavelet coefficients calculated by wavelet transform is as follows. First, the wavelet coefficient is obtained from the inner product of the captured image of the concrete surface and the wavelet function. By converting this wavelet coefficient to 256 gradations, a wavelet image having a continuous amount can be created.

  As described above, the wavelet coefficient varies depending on the crack width, crack density, and concrete surface density. As a result, the wavelet coefficient related to the crack density and concrete surface density for each gradation is created using simulated data. The wavelet coefficient table is created in advance. The wavelet coefficient for each gradation in the wavelet coefficient table serves as a threshold for crack detection. For example, wavelet coefficients (thresholds) corresponding to two contrasted concentrations (one concentration can be assumed to be a concrete surface concentration and the other concentration to be a crack concentration) are unique if the wavelet coefficient table is referenced. To be determined. Therefore, as will be described later, when the wavelet coefficient between the two densities to be compared in the captured image is calculated, if this wavelet coefficient is larger than the threshold value of the wavelet coefficient table, it can be determined that it is a crack, and the threshold value If it is smaller than that, it can be determined that it is not cracked.

  The pseudo data when creating the wavelet coefficient table is not particularly limited. For example, the crack width is 1 pixel (1 pixel) to 5 pixels (5 pixels), and cracks of each pixel width are included. For each, the wavelet coefficient corresponding to the gradation of the concrete surface and the gradation of the crack is calculated. When setting the threshold value, for example, the wavelet coefficient corresponding to the crack is selected from the wavelet coefficients when the crack width is 1 pixel, and the wavelet coefficient of the portion that is not the crack area is selected among the wavelet coefficients when the crack width is 5 pixels. A coefficient is selected, and an average value of these two wavelet coefficients can be used as a threshold value in an arbitrary gradation.

  In the crack detection method of the present invention, a route for creating a crack discrimination image obtained by removing minute noise (Method A), and a route for creating a non-crack discrimination image relating to surface noise such as a mold trace (Method B). One of the main features of the wavelet function applied to each creation route is to use a Gabor function suitable for detecting linear objects and a Laplacian Gaussian function suitable for detecting planar objects. .

  In either of the method A for creating a crack discrimination image by applying a Gabor function and the method B for creating a non-crack discrimination image by applying a Laplacian Gaussian function, the wavelet coefficient table described above is used in the steps 1A and 1B. A wavelet image is created by performing correction processing such as lens aberration correction and tilt correction on the photographed image, inputting it into a computer as an input image, and wavelet transforming the input image. The wavelet image is created in the computer as follows. First, wavelet coefficients are calculated for an appropriately set wide area (for example, an area of 30 × 30 pixels). Next, it is the same in a wide area moved by one pixel from this wide area (in the same way, for example, a 30 × 30 pixel area and most of the pixels are the same as the 30 × 30 pixel area before the movement). The wavelet coefficient is calculated as follows. By repeating this operation for the entire input image, a wavelet image composed of continuous amounts of wavelet coefficients is created inside the computer.

  Next, in Steps 2A and 2B of Method A and Method B, in the wavelet image composed of the continuous amount of the wavelet coefficients, the threshold value (wavelet coefficient) in the wavelet coefficient table is compared with the wavelet coefficients constituting the wavelet image. If the wavelet coefficient that constitutes is larger than the threshold value, it is determined as a crack or a planar object (for example, white on the screen), and if it is smaller than the threshold value, it is determined that it is not a crack or a planar object (for example, on the screen) Black). Then, by performing this operation on the entire wavelet image, a binary image in the middle stage in which a white crack or a planar object is drawn in a black background color is created.

  In Steps 2A and 2B in the method A and the method B, the contour line tracking process is performed on the binarized image in the middle stage, and the noise removal process is performed on the cracked or planar object.

  This contour line tracking process is to remove noise from cracks or planar objects. Contour tracking processing starts from an arbitrary pixel (a pixel that is determined to be a crack), and connects to the start pixel if the adjacent pixel is a cracked location, and further if the adjacent pixel is a cracked location Are connected to each other and finally closed to the starting pixel (for example, the first pixel, the second pixel,..., The n−1th pixel, the nth pixel, and the first pixel are connected in this order), The process ends when there is no longer any crack connected to the next. According to this contour line tracking process, an appropriate crack line such as a crack line that closes in a loop shape or a crack line that has a plurality of bent portions and extends linearly is created. At this time, by setting the minimum number of pixels to be connected in advance, it is possible to delete all pixels equal to or less than this set number from the crack display on the screen, assuming that they are not cracked.

  Next, in Steps 3A and 3B in Method A and Method B, the binarized images created in Steps 2A and 2B are cracked from the shape of a large number of objects (objects) formed by connecting pixels, and the formwork Discriminate traces, blotches, and artificial structures.

Specifically, in steps 3A and 3B of method A and method B, the feature amount of the lump is specified. Here, the feature amount of a lump is a factor that affects the shape of the lump. For example, the lump area, the height and width of the lump's circumscribed rectangle, and the major axis of the oval corresponding to the lump's outline And the minor axis, the ratio of the area of the lump and its circumscribed rectangle, the ratio of the width and height of the circumscribed rectangle of the lump, the ratio of the major and minor axes of the oval corresponding to the outer shape of the lump, the roundness of the lump ( (Perimeter length) ² ÷ (4π × area) The roundness of the lump calculated from the general formula, the irregularity of the lump outer circumference (fractal dimension), the length of the lump outer periphery, and the like.

  Regarding the feature quantities of these blocks, a feature quantity table is created in advance. The feature table may be created on a concrete surface that is currently subject to crack detection on the concrete surface, or on any other concrete surface. Step 1A, Step 2A, Step 1B, and Step 2B are executed in the same manner to create a binarized image, and a plurality of feature values of each block constituting the binarized image are calculated by a computer and tabulated. deep.

  The feature table further includes observation results regarding this arbitrary concrete surface, that is, the classification “3” when each lump is cracked, and the classification “2” when the noise type is a formwork or a scar. ”In the case of noise such as stains and dirt, among the noise types, the classification“ 1 ”, etc., and so on, whether each lump as an observation result belongs to a group of cracks or noise types Assigned.

  Discriminant analysis is performed on this feature quantity table with a plurality of feature quantities of each block as explanatory variables, and cracks or one of the noise types as an objective variable. It is determined that it is one of the noise types, and the determined result is used as a predicted value.

  For example, if the predicted value for a mass that is actually cracked is cracked, the predicted value matches the observed result, and the predicted value for a mass that is not actually cracked (a type of noise type) is cracked. If so, the predicted value and the observation result do not match.

  Here, “discriminant analysis” is a standard for discriminating which group is included when new data is obtained when it is clear that the pre-given data is divided into different groups. Is a technique for obtaining a discriminant function.

  In the method A, a plurality of feature amounts of each lump are used as explanatory variables, and cracks or one of noise types is used as an objective variable.

  Similarly, in the method B, a plurality of feature values of each lump is used as an explanatory variable, and peeling and peeling, a spot, an artificial structure, or any noise type is used as a target variable.

  In steps 3A and 3B, a crack discrimination image is obtained by performing identification and discriminant analysis of the feature amount for each of a large number of chunks incorporated in the binarized image specified in steps 2A and 2B. It is possible to create a planar non-crack discrimination image such as a mold trace or the like.

  When a Gabor function is applied in Method A to create a crack discrimination image and a Laplacian Gaussian function is applied in Method B to create a non-crack discrimination image, the difference between the crack discrimination image and the non-crack discrimination image in the computer is the last. By creating an image, a crack image corresponding to the difference image is created.

  According to the crack detection method of the present invention, a route for creating a crack discrimination image while removing fine noise by applying a Gabor function, and a non-related noise related to planar noise such as form marks by applying a Laplacian Gaussian function. By creating a crack image based on the difference between the crack identification image and the non-crack identification image created by each creation route, and creating a crack discrimination image, linear fine fine noise and formwork traces are created. Thus, it is possible to create a highly accurate crack image in which both surface noises such as the above are effectively removed.

  When the removal of noise has been completed and a crack image has been created, the crack width should be estimated. Here, an embodiment of the crack width estimation method will be described. When estimating the crack width, unlike the concrete surface to be photographed this time, it is a photographed image of the crack scale that has been pasted on the concrete surface of the photographed object, and the crack scale already accumulated. A photographed image is set as a database. The crack scale photographed image database includes a plurality of actual size values q from the minimum width to the maximum width of the crack scale, and further from a plurality of image data having different photographing distances, lens focal lengths, and spatial resolutions. It is composed.

  On the other hand, the crack image created in the third step on the concrete surface of the crack identification target is further thinned to form the center line of the cracked thin line of the concrete surface to be photographed. Create an image.

  The crack scale photographed image in the database is input to the computer as a crack scale input image, and the crack scale wavelet image is created by wavelet transform of the crack scale input image. The crack scale thinned image composed of the center line is created by executing the processing, and in the crack scale thinned image, the explanatory variable p of the crack width estimation formula is set to p = (wavelet coefficient value) ) − (Threshold of wavelet coefficient), and regression analysis is performed from the actual size value q of the crack scale specified from the captured image of the crack scale and the explanatory variable p. Then, the following crack width estimation formula with the explanatory variable p as a parameter: y = a + bp (a and b are constants specified by regression analysis, p is an explanatory variable) is obtained, and the object to be photographed with respect to the estimation formula p By applying p specified from the value of the wavelet coefficient in the thinned image of the crack on the concrete surface and the threshold value of the wavelet coefficient, the crack width of each concrete surface is specified for each subject.

  Here, the crack width estimation formula specified by the regression analysis varies depending on the amount of data of the accumulated crack scale and the like. The explanatory variable p is a value corresponding to the total number of pixels of the crack scale thinned image.

  As a result of the regression analysis, it is possible to create a database composed of crack width at crack scale, wavelet coefficient, luminance of target pixel, average luminance of surrounding pixels, spatial resolution and correlation coefficient.

  In addition, as a result of the regression analysis, the crack width estimation formula may be obtained using only a photograph image of a crack scale having a correlation coefficient of a certain height or higher. The accuracy of the estimation formula largely depends on the correlation coefficient obtained by the regression analysis. More specifically, the regression analysis is performed using only the photographed image of the crack scale having a certain high correlation coefficient. It is desirable to specify an estimation formula for the crack width. As a result, it is possible to remove a blurred image that depends on a problem of focus when the crack scale is photographed or a sticking method of the crack scale is floating, and as a result, the accuracy of the estimation formula is improved.

  Finally, crack data based on the estimated crack width, that is, calculation of the total crack length, calculation of crack length for each crack width, creation of a distribution map for each crack width, creation of a crack development map, etc. Desired information can be extracted from the crack data.

  Further, in a preferred embodiment of the crack detection method according to the present invention, the distance between each lump and the position of the center of gravity of the group in the feature quantity table is specified by the Mahalanobis general distance with respect to the feature quantity table. The probability value that a lump belongs to each group is specified from the Mahalanobis's general distance, and the determination result of each lump is corrected based on the probability value.

  The discriminant functions include a linear discriminant function based on a hyperplane / straight line and a non-linear discriminant function based on a Mahalanobis general distance based on a hypersurface / curve. In this embodiment, the distance between each block and the center of gravity position of the group in the feature table is specified by this Mahalanobis general distance, and the probability value that the block belongs to each group is specified from this Mahalanobis general distance. Then, the determination result of each lump is corrected based on this probability value.

  In this way, by correcting the determination result of each block based on the probability value, the predicted value for each block can be corrected to a corrected predicted value, and a crack image based on the corrected predicted value can be created. The detection accuracy is further improved.

  And regarding this probability value, a threshold value of the probability value is defined in advance, and if the specified probability value falls below this threshold value, it does not belong to any group, and for a mass that does not belong to any group For example, the group of the chunks may be specified artificially.

  In other words, for those that do not belong to any group, it is more efficient for the observer (engineer) to make a judgment based on the captured image and the probability value, rather than applying some algorithm to the computer. This is because the accuracy of the results is favorable.

  As can be understood from the above description, according to the crack detection method of the present invention, a route for creating a crack discrimination image while removing fine noise by applying a Gabor function, and a formwork by applying a Laplacian Gaussian function. By creating a crack image based on the difference between the crack discrimination image and the non-crack discrimination image created by each creation route, and creating a crack image for each line noise Thus, it is possible to create a highly accurate crack image in which both fine fine noise and surface noise such as formwork traces are effectively removed.

(A) is the schematic diagram which showed the relationship between an input image and a local region, (b) is the schematic diagram which showed the relationship between a local region and an attention pixel. It is the flow which showed the crack detection method of the present invention. It is the figure which showed the pseudo image. FIG. 4 is a bird's-eye view of wavelet coefficients of the pseudo image of FIG. 3. It is the figure which showed one Embodiment of the wavelet coefficient table. It is a flowchart of embodiment of the calculation method of the estimated value of the crack width in the flow of FIG. It is a flowchart of other embodiment of the calculation method of the estimated value of a crack width. (A) is the photograph figure which showed the picked-up image, (b) is the figure which showed the binarized image created by the method A. (A) is the figure which showed the crack discrimination | determination image produced by the method A, (b) is the figure which showed the binarized image produced by the method B. (A) is the figure which showed the non-crack discrimination | determination image (overhead detection image) created by the method B, (b) showed the non-crack discrimination image (black spot detection image) created by the method B FIG. It is the figure which showed the crack image produced based on the difference of the crack discrimination | determination image of FIG. 9a, and the non-crack discrimination | determination image of FIG.

  Hereinafter, embodiments of a crack detection method of the present invention will be described with reference to the drawings. FIG. 1 a is a schematic diagram showing a relationship between an input image and a local region. In the crack detection method of the present invention, wavelet transform is performed in the wide area 2 in the input image 1, and cracks are detected in the local area 3 that is the center of the wide area 2. The wide area 2 is translated vertically and horizontally throughout the input image 1 to detect cracks in the input image 1. By this method, it is possible to detect cracks with higher accuracy as compared with the method of determining one threshold value in the input image 1 as in the conventional fixed threshold method.

  FIG. 1 b is an enlarged view of the local region 3. In the illustrated embodiment, for example, nine pixels of 3 × 3 (eight neighboring pixels 31, 31... And a target pixel 32 located in the center) are targeted. As a crack judgment. The wavelet coefficient is calculated for the wide area 2 in FIG. 1a. Here, in method A in the flow shown in FIG. 2, wavelet coefficients are calculated by performing wavelet transform using the following equation using a Gabor function as a wavelet function (mother wavelet function).

[Equation 1]

[Equation 2]

[Equation 3]

Here, f (x, y) is an input image (where x and y are arbitrary coordinates in the two-dimensional input image), Ψ is a mother wavelet function (Gabor function), and (x ₀ , y ₀ ) is the translation amount of Ψ, a ^k is the expansion or contraction of Ψ (where a ^k is the reciprocal of the frequency, and the frequency width for calculating several frequency regions is indicated by an integer k) value), rotate the f ₀ is the center frequency, the σ is the standard deviation of the Gaussian function, the rotation angle theta is representative of the direction of travel of the wave, (x ', y') is the angle theta (x, y) Each coordinate is shown.

Here, the cumulative value C (x ₀ , y ₀ ) of the wavelet coefficient Ψ is obtained for a plurality of θ, k calculated using Formula 1 as Formula 4.

[Equation 4]

The above parameters can be set arbitrarily. For example, σ can be set to 0.5 to 2, a ^k to 0 to 5, f ₀ to 0.1, and the rotation angle to 0 to 180 degrees. .

The parallel movement amount (x ₀ , y ₀ ) in Equation 4 corresponds to the position of the target pixel, and the continuous amount of wavelet coefficients (C (x ₀ , y ₀ ) is obtained by sequentially moving the position of the target pixel. )) Can be calculated, and a wavelet image can be created by illustrating this continuous quantity.

  On the other hand, in the method B in the flow shown in FIG. 2, wavelet coefficients are calculated by performing wavelet transformation using the following equation using a Laplacian Gaussian function as a wavelet function (mother wavelet function). Hereinafter, a Laplacian Gaussian function is expressed by Equation 5, a wavelet expansion coefficient (two-dimensional wavelet transform basic equation) is expressed by Equation 6, and a feature amount including a cumulative value of absolute values of the wavelet expansion coefficients expressed by Equation 6 is expressed by Equation 7, respectively. .

[Equation 5]

[Equation 6]

[Equation 7]

Here, f (x, y) represents an input image, (x ₀ , y ₀ ) represents a translation amount on the plane (x, y), and σ represents a standard deviation of a Gaussian function.

  For all the pixels that make up the wide area 2, both method A and method B can calculate the wavelet coefficient based on the unique upper calculation formula, and then move the target pixel up and down or up and down. Wavelet coefficients are calculated in the same manner for all the pixels in region 2. By performing this wavelet coefficient calculation on the entire input image, it is possible to create a wavelet image (an image made up of a continuous amount of wavelet coefficients) in which the constituent pixels within an appropriate range have respective wavelet coefficients.

  Next, an embodiment of a crack detection method will be described with reference to FIG.

First, an input image is created (step S10) by taking a photographed image of a concrete surface photographed by a digital camera such as a CCD camera into a computer.

  Next, in both method A and method B, wavelet coefficients are calculated for a pseudo image having two contrasts that have no relationship with the input image. For example, as shown in FIG. 3, the background color a assumed to be a concrete surface (for example, the background colors R, G, B are assumed to be 255, 255, 255) and the line segments b1 to b1 assumed to be cracks. The wavelet coefficient of the pseudo image consisting of b5 is obtained. Here, the line segments b1 to b5 change in order from 1 pixel (1 pixel) to 5 pixels (5 pixels) in order, and each line segment has three types of density (for example, In the line segment b1, it changes to b11 (black), b12 (light black), and b13 (gray) in descending order of density). FIG. 4 shows a bird's-eye view of wavelet coefficients calculated by performing wavelet transform on this pseudo image. In FIG. 4, the X axis indicates the line segment width, the Y axis indicates the color density of the line segment, and the Z axis indicates the wavelet coefficient. The line width may be set according to the maximum crack width to be finally extracted. For each pixel width, the wavelet coefficient of the crack region and the wavelet coefficient other than the crack region can be calculated. In this embodiment, when calculating the arbitrary density (gradation) assumed to be a concrete surface and the threshold value (wavelet coefficient) corresponding to the arbitrary density (gradation) assumed to be a crack, for example, the crack width is Threshold value corresponding to the target gradation within the set crack width range, with the average value of the wavelet coefficient of the crack area in the case of 1 pixel width and the wavelet coefficient other than the crack area in the case of the crack width of 5 pixels. It is said. Of course, this threshold value may be arbitrarily set.

  By performing the two density combinations to be compared with 256 gradations of 0 to 255, the wavelet coefficient table as shown in FIG. 5 is created (steps S30A and 30B in FIG. 2). Note that this operation does not have to be performed at the illustrated flow position, and may be performed, for example, before creation of the input image.

  Returning to FIG. 2, in method A, wavelet transform is applied to the input image by applying a Gabor function, and in method B, wavelet transform is performed by applying the Laplacian Gaussian function to the input image (step S20A). , S20B). A wavelet image is composed of a continuous amount in which each pixel has a unique wavelet coefficient as described above. By comparing the wavelet coefficient of each pixel with the wavelet coefficient (threshold) of the corresponding wavelet coefficient table, the wavelet image is black. A binarized image in the middle stage in which white cracks are drawn in the background color (binarized image from which a part of noise is not removed) is created.

  Next, the outline of the outline tracking process will be outlined. This contour tracking process starts from an arbitrary cracked pixel in each cracked area (first pixel), for example, pays attention to a pixel adjacent counterclockwise from this first pixel, and this adjacent pixel (second pixel) If is a cracked pixel, the first pixel and the second pixel are connected. Thereafter, similarly, the second pixel, the third pixel,..., The (n−1) th pixel, the nth pixel and the cracked pixel are traced, and when the first pixel that is the starting point comes after the nth pixel, One pixel to n-th pixel are determined as one crack location (crack line). Alternatively, when there is no crack pixel following the nth pixel, the first pixel to the nth pixel are determined as one crack location (crack line). The crack line also includes a crack that branches into two or more branches in the middle. Setting of the order n is arbitrary, and a binarized image in which a part of noise is removed is created by determining that the tracking number from the first pixel is equal to or more than the set order n as a crack. Is done.

  When a binarized image is created by the above method, image editing software is used on the binarized image to remove the dot portion, and a line segment less than a predetermined length is removed as a non-cracked portion. By applying such a method, the noise is finally removed to create a crack discrimination image (step S50A).

  In the methods A and B, a feature amount table based on the binarized image is created independently of the processing flow up to the creation of the binarized image (steps S40A and S40B).

Specifically, a binarized image is created by executing steps S40A and S40B in FIG. 2 on a concrete surface different from the concrete surface to be cracked. This binarized image contains a large number of chunks (objects) formed by connecting pixels having the same wavelet function, and some feature quantities of each chunk are identified and stored in the feature quantity table. As a table. Here, the feature amount of the lump is a factor that affects the shape of the lump, and the area of the lump, the height and width of the circumscribed rectangle of the lump, the major and minor axes of the ellipse corresponding to the outer shape of the lump, The ratio of the area of the lump to its surrounding rectangle, the ratio of the width and height of the lump's circumscribed rectangle, the ratio of the major axis to the minor axis of the oval corresponding to the outer shape of the lump, the roundness of the lump ((perimeter length) ² ÷ (4π × area) The roundness of the lump calculated from the general formula), the roughness of the lump's outer periphery (fractal dimension), the length of the lump's outer periphery, etc.

  And the observation result regarding this arbitrary concrete surface is further input into this feature-value table. Specifically, when the lump is cracked, the classification is “3”, among the noise types, the classification is “2”, and among the noise types, the classification is “stains or dirt”. A number such as “1” is defined, and it is assigned to which of the groups of cracks and noise species each lump as an observation result belongs (step S60).

  Next, with reference to this feature amount table, a discriminant analysis is performed on a large number of masses contained in the binarized image created for the concrete surface to be cracked, and a predicted value is obtained. . More specifically, a discriminant analysis is performed by using a plurality of feature quantities of each lump as explanatory variables and cracking or any noise type as an objective variable. It is determined that it is one of the noise types, and the determined result is set as a predicted value (step S61).

  Prediction values based on the discriminant analysis of a large number of objects (objects) built in the binarized image with respect to the concrete surface to be cracked are added to the feature value table created in step S60, and the feature value table becomes It will be updated sequentially. Note that the actual crack observation results for the concrete surface subject to crack identification are also input to the feature quantity table.

  The predicted value of each block is specified by discriminant analysis and automatically input to the feature table in the computer, but the remaining noise is the one whose predicted value does not match the observed value.

  This noise is corrected using the probability value (step S63, the calculation of the probability value is step S62), the predicted value is corrected and the feature amount table is updated, and this content is reflected in the binarized images S40A and S40B. Then, the remaining noise is removed to create a binarized image (steps S40A and S40B).

  Here, an outline of a method for correcting a predicted value by applying a probability value will be described. In the feature amount table (not shown), the number of the predicted values that have become the determination categories (for example, “1” to “3”) with respect to the observation value category (for example, “1”) is represented. For example, the number of positive discriminations in which the observed value category “1” becomes the predicted value judgment category “1” is measured, and the number of misclassifications in which the observed value category “1” becomes the predicted value judgment category “2”. And the number of misclassifications in which the observed value category “1” becomes the predicted value judgment category “3” is measured. Further, the correct discrimination rate and the incorrect discrimination rate are obtained.

  As a consideration on the results of discriminant analysis, the observed value category “3” → the predicted value judgment category “3” is measured, the observed value category “3” → the predicted value judgment category “2” is measured, and the observed value Category “3” → predicted value judgment category “1” is measured. Here, the reason why the observed value classification “3” → predicted value judgment classification “1” or “2” is the case where cracks and cracks, or spots such as spots are combined, or crack detection is insufficient. This is because it becomes discontinuous, and the misclassification rate tends to be high for objects in a partial area of the crack. Also, the observed value category “1” → the predicted value judgment category “1” is measured, the observed value category “1” → the predicted value judgment category “2” is measured, and the observed value category “1” → The judgment classification “3” of the predicted value is measured. Also, the observed value category “2” → the predicted value judgment category “1” is measured, the observed value category “2” → the predicted value judgment category “2” is measured, and the observed value category “2” → The judgment classification “3” of the predicted value is measured. The misclassification rate of the observed value category “1” → predicted value judgment category “2” or “3” is extremely rare, but may occur due to dirt or patterns on the concrete surface. Misclassification of observed value category “2” → predicted value judgment category “1” or “3” may occur in part of spot detection, form trace, and crack detection on the concrete surface.

  In response to such a determination result, the predicted value is corrected by applying the probability value. When the observed value classification is “3” → the predicted value judgment classification is “1” or “2”, there is a need for improvement in order to erroneously determine that the crack is not a crack. Therefore, it is determined which group each object (object) belongs to based on the Mahalanobis general distance. Since it is not clear how much the group belongs to this method, the probability that each lump belongs to each group is calculated using Mahalanobis's general distance. As for the Mahalanobis generalized distance, we referred to “Multivariate Analysis Method (Tadakazu Okuno et al.) Nisshingi Renren”.

Relates calculation of the probability values using Mahalanobis distance, the method of discrimination in discriminant analysis a new mass (object _chi) is determined by the distance between the position of the center of gravity G _h for each group grouped by the feature amount data. In the discriminant analysis, the distance d (χ, G _h ) between the object χ and the centroid position G _{h of} each group, h = 1, 2,... G is used, and the Mahalanobis general distance is used. Here, g represents the number of groups, and p represents the dimension of the object χ. When χ = {x ₁ , x ₂ ,... x _p } follows a p-dimensional normal distribution N (μ, Σ), its mean vector μ and population variance / covariance matrix Σ for a group are given by It can be written as 8.

[Equation 8]

[Equation 9]

  Further, the probability density of the p-dimensional normal distribution can be written as the following Expression 10.

[Equation 10]

Rewriting Equation 10 by using the Mahalanobis distance D ² can be written as the following equation 11.

[Equation 11]

D ² defined by Equation 9 follows a χ ² distribution with p degrees of freedom. Probability is larger than a certain value D ² can be determined using the chi ² distribution.

  In order to reduce the misclassification of the observation value classification “3” → the prediction value judgment classification “1” or “2” by examining the misclassification rate of the feature value data using the probability value thus obtained. The predicted value correction method will be described below.

  The threshold value is defined as a constant value for the probability value, and if the probability of belonging to each group of an object is less than or equal to the threshold value, a discrimination category “others” is added to the predicted value and it belongs to this discrimination category. To do. An engineer determines an object belonging to this discrimination category “others” based on a captured image and a probability value.

  Based on the discriminant analysis result based on the feature value data after the predicted value correction, the handling of objects that do not belong to any group of the predicted value determination categories “1 to 3” was examined. When the predicted value does not belong to any group of the determination categories “1 to 3”, the determination category “others” is added to the predicted value, and when the object enters the determination category “others”, the predicted value is calculated from the probability value. Calculate the correction value.

  Regarding this corrected predicted value, an engineer will make a final judgment with reference to the descending order of probability values.

  In the conventional method, since noise other than cracks created in the step of creating a crack discrimination image in step S50A in FIG. 2 is performed manually using image editing software, this operation is extremely complicated and time-consuming. Therefore, human error such as oversight of cracks was likely to occur during the operation process. On the other hand, the crack discriminating image and the non-cracking discriminating image in steps S50A and S50B obtained by performing discriminant analysis and automatically processing the crack determination while correcting the predicted crack value using the probability value. Such an artificial mistake is eliminated by the cracked image (step S70) obtained by the difference.

  Next, the crack discrimination image created by Method A and the non-crack discrimination image created by Method B are combined in a computer, and more specifically, the difference processing between the crack discrimination image and the non-crack discrimination image is performed. By creating an image, the planar non-crack discrimination image included in the crack discrimination image is removed to create a crack image (step S70).

  In this way, according to the processing up to step S70 in the flow of FIG. 2, a route (method A) for creating a crack discrimination image while applying a Gabor function to remove fine noise and a Laplacian Gaussian function are applied. The route (Method B) for creating a non-crack discriminating image for surface noise such as mold traces is performed in parallel, and the crack is determined by the difference between the crack discriminating image and the non-cracking discriminating image created by each creation route. By creating an image, it is possible to create a high-accuracy cracked image in which both linear and fine fine noise and surface noise such as formwork traces are effectively removed. Here, “performing method A and method B in parallel” only means that there are two routes, and two routes may be performed simultaneously. Method B may be executed after execution, and does not impose time constraints on the execution of both methods.

  When a crack image is created in step S70, a thinned image is formed with the center line of each crack for the created crack image, for example, the entire crack has one pixel width (one pixel width). Then, an estimated value of the crack width is calculated using the thinned image (step S80). Here, the estimation method of this crack width is demonstrated based on FIG.

  The crack scale wavelet image is generated by inputting the photographed image of each crack scale stored in the database in step S100 to a computer to create an input image of each crack scale, and wavelet transforming the input image of each crack scale. (Step S101), and the wavelet coefficient of each pixel of the wavelet image is compared with the wavelet coefficient (threshold value) of the corresponding wavelet coefficient table to create a binary image of the crack scale (step S102). Then, thinning processing is executed on this to create a crack scale thinned image composed of the center line (step S103).

  In the created crack scale thinned image, the explanatory variable p of the crack width estimation formula is defined by the following formula 12.

[Equation 12]
p = (wavelet coefficient value) − (wavelet coefficient threshold)

  Regression analysis is performed from the explanatory variable p described above and the actual size value q of the crack scale specified from the captured image of the crack scale.

  As a result of the regression analysis, the following crack width estimation formula using the explanatory variable p as a parameter can be obtained (step S104).

[Equation 13]
y = a + bp
(Unit is mm, a, b are constants specified by regression analysis, p is explanatory variable)

  By applying p specified from the wavelet coefficient value and the wavelet coefficient threshold value in the thinned image of cracks on the concrete surface to be photographed to p in the estimation formula, the crack width of each concrete surface for each object to be photographed Can be specified (step S104).

  Returning to FIG. 2, after estimating the crack width in step S80, a table and drawings (crack development etc.) relating to the distribution state of each crack, the number of cracks for each crack width, etc. are created as data (step S90).

  In FIG. 6, the flow from the crack scale database to the identification of the crack width estimation formula and the creation of a thinned image of the concrete surface to be imaged are shown in a parallel flow diagram. The method is not limited to this flowchart. For example, steps S100 to S104 are performed in advance, an estimation formula for the crack width is specified in advance, and then the thinning on the concrete surface to be photographed is performed. It may be a flow for creating an image.

  Here, as a result of the regression analysis, it is desirable to perform a regression analysis using only a crack scale image having a relatively high correlation coefficient to specify an estimation formula for a crack width.

  For example, regression analysis is performed using a crack scale image having a correlation coefficient of 0.85 or more, and the specified crack width estimation formula (for example, y = 0.029 + 0.000736p) can identify the crack width with extremely high accuracy. it can.

  FIG. 7 is a diagram illustrating another embodiment relating to a crack width estimation method.

  The p obtained from the captured image tends to vary depending on the quality of the captured image. Therefore, in order to standardize the quality of the captured image with a certain standard, the standardized p represented by the following equation is obtained.

[Equation 14]

[Equation 15]

[Example]
The inventors actually performed the crack detection method of the present invention. FIG. 8A is a photographic view showing a captured image, and FIG. 8B is a view showing a binarized image created by the method A. 9A is a diagram showing a crack discrimination image created by the method A, and FIG. 9B is a diagram showing a binarized image created by the method B. FIG. 10a is a diagram showing a non-crack discrimination image (overhead detection image) created by Method B, and FIG. 10b is a non-crack discrimination image (black spot detection image) created by Method B. FIG. Further, FIG. 11 is a diagram showing a crack image created based on the difference between the crack determination image of FIG. 9a and the non-crack determination image of FIGS. 10a and 10b.

  According to the verification by the present inventors, regarding the crack discrimination image shown in FIG. 9a, the discrimination rate between actual cracks and noise such as dirt on the concrete surface is as high as 92% or more, and discrimination between actual cracks and form marks Since the rate is not as high as 70%, only noise removal was performed.

  Therefore, discriminant analysis is performed on each object of the binarized image of FIG. 9B obtained by applying the Laplacian Gaussian function to the captured image, and the overhead line detection image shown in FIG. 10A, the black spot detection image shown in FIG. 10B, After obtaining the mold trace detection image and the like, the crack image shown in FIG. 11 was created by automatic processing based on the difference between the crack discrimination image shown in FIG. 9a and the non-crack discrimination image shown in FIGS. According to this cracked image, it is a highly accurate cracked image from which formwork marks and the like are effectively removed.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and there are design changes and the like without departing from the scope of the present invention. However, they are included in the present invention.

DESCRIPTION OF SYMBOLS 1 ... Input image, 2 ... Wide area | region, 3 ... Local area | region, 31 ... Neighboring pixel, 32 ... Interesting pixel

Claims (2)

(Method A) A method of creating a crack discrimination image for detecting cracks occurring on the concrete surface, the method comprising Step 1A, Step 2A, and Step 3A,
(Method B) Plane objects represented by artificial structures such as stains, springs, peeling and peeling, and artificial structures such as cables and overhead lines, which are objects having a spread on the surface of the concrete. A non-crack discrimination image corresponding to the method, comprising: Step 1B, Step 2B, and Step 3B;
A crack detection method comprising (Method A) and (Method B),
(Method A)
Simulate the density of the crack and the concrete surface, apply the Gabor function and calculate the wavelet coefficients corresponding to the two concentrations to be compared, and the wavelet coefficients when these two concentrations are changed respectively. Step 1A for creating a wavelet coefficient table by calculating a wavelet coefficient table, inputting a captured image of a concrete surface to be cracked into a computer as an input image, and wavelet transforming the input image;
In the wavelet coefficient table, the average density of neighboring pixels in the local area assumed to be the density of the concrete surface and the wavelet coefficient corresponding to the density of the target pixel assumed to be the crack density are used as threshold values, and in any neighboring pixels When the wavelet coefficient of an arbitrary pixel of interest is larger than a threshold value, the pixel of interest in the neighboring pixel is determined as a cracked pixel, and the wavelet coefficient of an arbitrary pixel of interest in an arbitrary neighboring pixel is smaller than the threshold value. determines that no cracking of the target pixel in the near neighbor pixel, compares the wavelet coefficients and the threshold value of the target pixel while varying the local region and the pixel of interest, not further such to put the contour tracing processing, connecting the cracking pixel a single lump formed, the threshold of mass the mass is set in advance If remote small, it is determined that noise, by if greater than the threshold value of the mass is determined that cracking, steps 2A to create a binarized image in which a part of the noise is being removed,
Identify the feature quantity of the lump, refer to the feature quantity table in which cracks and noise types other than cracks are defined in advance for each feature quantity of the lump, use multiple feature quantities as explanatory variables, Perform a discriminant analysis with the noise type as the objective variable to determine that the mass with the specified feature quantity is either a crack or a noise type,
Step 3A, which creates a crack discriminating image in which noise is further removed from the binarized image by executing the identification of the feature amount for each chunk and discriminating and analyzing each chunk in Step 2A. ,
The feature amount table is created by executing steps 1A and 2A on an arbitrary concrete surface to create a binarized image, and a plurality of feature amounts of the respective blocks constituting the binarized image. It is calculated by computer and assigns one kind of group consisting of cracks and noise species, which is the observation result for each lump,
(Method B)
When the density of the object and the density of the concrete surface are set in a pseudo manner, the wavelet coefficients corresponding to the two densities to be compared are calculated by applying the Laplacian Gaussian function, and the two densities are changed respectively. A wavelet coefficient table is calculated and a wavelet coefficient table is created, and the captured image of the concrete surface that is the detection target of these objects is input to a computer as an input image, and a wavelet image is created by wavelet transforming this input image Step 1B,
In the wavelet coefficient table, the average density of neighboring pixels in the local region assumed to be the density of the concrete surface, and the wavelet coefficient corresponding to the density of the target pixel assumed to be the density of the planar object are set as threshold values, and any arbitrary When the wavelet coefficient of an arbitrary pixel of interest in a neighboring pixel is larger than a threshold, the pixel of interest in the neighboring pixel is determined as a planar object, and the wavelet coefficient of an arbitrary pixel of interest in an arbitrary neighboring pixel is smaller than the threshold In this case, it is determined that the target pixel in the neighboring pixel is not a planar object, the wavelet coefficient of the target pixel is compared with the threshold value while changing the local region and the target pixel, and the pixel determined as the planar object It performs determination processing of the planar object and noise to further perform the contour tracing process, the planar target And one mass is formed by connecting the determined pixel, if smaller than the threshold of mass the mass is set in advance is determined to noise, if greater than the threshold mass and the planar object Step 2B for creating a binarized image in which a part of noise is removed by determination,
Identify the feature quantity of the lump, refer to the feature quantity table that prescribes for each feature quantity of the lump, deformations such as spots, spring water and peeling off, artificial structures such as cables and overhead lines, etc. , Using a plurality of feature quantities as explanatory variables, performing a discriminant analysis with these planar objects as objective variables, and determining that the mass in which the feature quantities are specified is a planar object,
From Step 3B, which creates a non-crack discrimination image in which noise is further removed from the binarized image by performing the identification of the feature amount for each chunk and the discrimination analysis of each chunk for all the chunks in Step 2B. Become
Next, a crack detection method for creating a crack image based on a difference between a crack discrimination image created by Method A and a non-crack discrimination image created by Method B.   For the feature amount table in step 3B, the distance between each block and the center of gravity position of the group in the feature table is specified by the Mahalanobis general distance, and the block is assigned to each group from the Mahalanobis general distance. The crack detection method according to claim 1, wherein a probability value to belong is specified, and a determination result of each lump is corrected based on the probability value.

Images