Dataset, machine learning model, and program for analyzing tissue image

Abstract

To suppress the loss of a part of a line representing a grain boundary, when creating a grain boundary image from a tissue image including grain boundaries using a machine learning model.SOLUTION: A dataset includes input data consisting of a tissue image including crystal grains and grain boundaries, and correct answer data consisting of a processed image including a band region in which the width of the grain boundary is expanded and a remaining background region. Pixel values of the band region gradually decrease from pixel Aout located at the outermost edge of pixel A corresponding to the grain boundary to pixel Bout located at the outermost edge of pixel B corresponding to the band region. The machine learning model is obtained by machine learning such a dataset using artificial intelligence. A program for analyzing tissue images is configured to cause a computer to execute a procedure of outputting an inferred image including the band region and the background region from the tissue image and performing binarization processing and thinning processing on the inferred image.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The present invention relates to a dataset, a machine learning model, and a program for analyzing tissue images, and more specifically, to a machine learning model capable of converting a linear grain boundary into a band region wider than the grain boundary, a dataset for generating such a machine learning model, and a program for analyzing tissue images including such a machine learning model. [Background technology]

[0002] The grain size of polycrystalline materials affects mechanical properties such as strength and toughness. In particular, the grain size of metallic materials has a large effect on mechanical properties. Therefore, accurate measurement of the grain size of polycrystalline materials is important for estimating the mechanical properties of polycrystalline materials. Known methods for measuring the grain size of metal materials include the intercept method, the comparison method, and the counting method. Of these, the intercept method involves drawing straight lines on a micrograph and determining the average length of the line segments that cross each crystal grain. The intercept method is widely used because it is relatively easy to correlate grain size with mechanical properties. However, the intercept method requires a great deal of effort and has problems such as large individual differences in measurement results.

[0003] On the other hand, there are also known methods for analyzing the structure of polycrystalline materials using image analysis software and machine learning models. Methods using image analysis software and machine learning models have the advantage of being able to process large amounts of data quickly.

[0004] For example, Patent Document 1 states: Including input data and correct answer data, The input data is 3-pixel x 3-pixel image data extracted from a structural image of a granular material in which the surfaces of metal particles (granular regions) are covered with an oxide film (non-granular regions), The correct answer data is data for recognizing the pixel value of the center pixel of a 3 pixel x 3 pixel image data (hereinafter also referred to as the "center pixel") as zero (black) when the center pixel corresponds to a metal particle, and for recognizing the pixel value of the center pixel as one (white) when the center pixel corresponds to an oxide film. The data set is disclosed.

[0005] The same document states: (A) When binarizing a structural image of a granular material, if the image data is binarized based on a threshold value, one metal particle may be recognized as multiple metal particles; and (B) When 3 pixel x 3 pixel image data is used as input data, the pixel value (zero or one) of the binarized central pixel is used as the correct answer data for machine learning, and the resulting machine learning model is used to binarize the structure image of a granular material, improving the accuracy of metal particle recognition. is stated.

[0006] The dataset described in Patent Document 1 is a dataset for machine learning of a structural image of a granular material containing metal particles (granular region) and an oxide film (non-granular region). Therefore, a machine learning model generated using the dataset described in Patent Document 1 is not suitable as a machine learning model for identifying grain boundaries contained in a structural image containing multiple crystal grains (for example, a γ single-phase structural image of nickel-based alloy A716).

[0007] On the other hand, when a tissue image is binarized using an image analysis device, a grain boundary image (binarized image) in which the grain boundaries are depicted as lines can be obtained. However, when binarization is performed using a conventional image analysis device, parts of the lines representing the grain boundaries are often missing. This is thought to be due to the presence of image noise or differences in shading in the lines representing the grain boundaries. Therefore, in order to accurately identify the grain boundaries, manual marking of the grain boundary image (binarized image) output from the image analysis device was essential.

[0008] Alternatively, when machine learning is performed using a structural image and a corresponding grain boundary image (binarized image) as training data, a machine learning model capable of converting a structural image into a grain boundary image is obtained. However, even when a grain boundary image is output using the machine learning model obtained in this way, some of the lines representing the grain boundaries may be missing. Therefore, when the number, grain size, grain size distribution, etc. of crystal grains contained in a structural image are evaluated using a conventional method, sufficient evaluation accuracy may not be obtained. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] Patent Publication No. 2021-018752 Summary of the Invention [Problem to be solved by the invention]

[0010] The problem to be solved by the present invention is to prevent a portion of a line representing a grain boundary from being missing when a grain boundary image is created from a structural image including grain boundaries using a machine learning model. Another problem to be solved by the present invention is to improve the evaluation accuracy when evaluating the number, grain size, grain size distribution, etc. of crystal grains based on grain boundary images created using a machine learning model.

[0011] Another problem that the present invention aims to solve is to provide a dataset from which such a machine learning model can be generated. Furthermore, another problem to be solved by the present invention is to provide a program for tissue image analysis equipped with such a machine learning model. [Means for solving the problem]

[0012] In order to solve the above problem, a data set according to a first embodiment of the present invention comprises: Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; Equipped with The pixel value of the band region is the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out It is gradually decreasing towards

[0013] The data set according to the second embodiment of the present invention is Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; Equipped with The pixel value of the band region is the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out is gradually increasing towards

[0014] The machine learning model according to the present invention is obtained by training an artificial intelligence to learn machine learning from the dataset according to the present invention.

[0015] Furthermore, the tissue image analysis program according to the present invention is configured to cause a computer to execute the following procedures. (A) Step A for storing a structural image including crystal grains and grain boundaries in a memory. (B) Step B of inputting the tissue image into a machine learning model according to the present invention that has been pre-stored in the memory, causing the machine learning model to output an inferred image including a band region and a background region, and storing the inferred image in the memory. (C) A step C of performing a binarization process on the inference image to obtain a binarized image including the binarized band region, and storing the binarized image in the memory. (D) A step D of performing a thinning process on the binary image to obtain a grain boundary image in which the band regions are converted into lines, and storing the grain boundary image in the memory. Effect of the Invention

[0016] A machine learning model is obtained by using a structural image including grain boundaries as input data and a processed image including a band region where the grain boundaries have been widened as correct answer data for machine learning. When an arbitrary structural image is input to the obtained machine learning model, an inferred image in which the linear grain boundaries have been converted into a wide band region is output. The obtained inferred image is subjected to a binarization process, and the binarized band region is subjected to a thinning process, resulting in a grain boundary image corresponding to the input structural image. The grain boundary image obtained in this way is less likely to lose parts of the lines representing the grain boundaries than grain boundary images obtained by the conventional method, which is believed to be because the grain boundaries are converted into wide band regions before being binarized.

[0017] Furthermore, by optimizing the pixel value gradient of the band region, the difference between the position of the line on the grain boundary image and the position of the original grain boundary can be reduced. This makes it possible to prevent large grains from being displayed smaller than their actual size in the grain boundary image. This also makes it possible to prevent a decrease in the evaluation accuracy of the number of grains, grain size, grain size distribution, etc. [Brief description of the drawings]

[0018] [Figure 1] FIG. 13 is a schematic diagram of a method for generating a processed image from a structure image including grain boundaries, in which the grain boundaries are converted into band regions wider than the grain boundaries. [Diagram 2] 1A and 1B are diagrams showing an example of a tissue image and a binarized image. [Diagram 3] 1A-1C are diagrams illustrating an example of a distance transformed image, a gradient increased image, and a gradient adjusted image. [Figure 4] FIG. 2 is a flow diagram of a tissue image analysis program according to the present invention. [Diagram 5]1A to 1C are diagrams showing examples of an inferred image, a binarized image, and a corrected grain boundary image created using the program for analyzing structural images according to the present invention.

[0019] [Figure 6] 1 shows an example of a dataset used in Comparative Example 1, and an example of an inference image created based on the machine learning model of Comparative Example 1. [Figure 7] 1 shows an example of a data set used in Example 1, and examples of an inferred image, a binarized image, and a grain boundary image created based on the machine learning model of Example 1. [Figure 8] 13 shows an example of a data set used in Example 2, and examples of an inferred image, a binarized image, and a grain boundary image created based on the machine learning model of Example 2. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] An embodiment of the present invention will be described in detail below. [1. Dataset (1)] The data set according to the first embodiment of the present invention is Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; It is equipped with: In this embodiment, the pixel value of the band region is determined based on the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out It is gradually decreasing towards

[0021] Hereinafter, the pixel corresponding to the grain boundary will be simply referred to as "pixel A." In addition, among the pixels A, the pixel located at the outermost edge of the grain boundary will be specifically referred to as "pixel A out "It is also called. The pixels corresponding to the band region other than pixel A are also simply referred to as "pixel B." In addition, among the pixels B, the pixels located at the outermost edge of the band region are specifically referred to as "pixel Bout "It is also called. Furthermore, the pixel corresponding to the background region is simply referred to as "pixel C." out The pixel C ad "It is also called.

[0022] [1.1. Tissue images] "Structural image" refers to an image that includes crystal grains and grain boundaries. A structural image is obtained by converting a micrograph of various materials that includes crystal grains and grain boundaries into digital data. The structural image is used as input data for machine learning. The material to which the present invention is applied is not particularly limited as long as it contains crystal grains and grain boundaries. Examples of the material to which the present invention is applied include nickel-based alloys, chromium-based stainless steels, structural steels, and NdFeB-based magnet alloys.

[0023] The tissue image may be a grayscale image or a color image. The number of gradations is not particularly limited as long as it is possible to form a gradient in the pixel values ​​of the pixels included in the band region. In the case of a grayscale image, the number of gradations is usually 256 gradations. The size of the tissue image is not particularly limited, and an optimal size can be selected depending on the purpose. For example, multiple large tissue images may be acquired and used as input data as is. Alternatively, one large tissue image may be divided into small regions and images of the multiple small regions may be used as input data.

[0024] [1.2. Processed images] [1.2.1. Definition] The term "processed image" refers to an image that includes a band region and a remaining background region. The processed image is obtained by performing a predetermined processing process on a tissue image. The processed image is used as the correct answer data when performing machine learning. In this embodiment, the "predetermined processing" refers to a process of adjusting the pixel values ​​and pixel value gradients of the tissue image so that grain boundaries are converted into band regions and the band regions are displayed brighter than the background region.

[0025] The "band region" refers to a region in which the width of the grain boundary included in the structure image is expanded toward the inside of two adjacent crystal grains. Among the pixels included in the band region according to the present embodiment, the pixels with the maximum pixel value are all of pixel A, and the pixels with the minimum pixel value are all of pixel B. out Furthermore, the pixel value of the band region is pixel A out From pixel B out It is gradually decreasing towards "Background region" refers to the region remaining after removing the band regions from the tissue image.

[0026] "Pixel value gradually decreases" means that pixel A out From pixel B out The band area is a region where the pixel values ​​decrease continuously or stepwise from A to A. out From B out The pixel value may decrease linearly or may decrease in a curved manner toward the pixel value.

[0027] [1.2.2. Pixel value] After generating a machine learning model using the data set according to the present embodiment, when an arbitrary tissue image is input to the machine learning model, an inference image in which grain boundaries are converted into band regions is output. The inference image is further subjected to binarization and thinning processes to become a grain boundary image (binarized image). Therefore, the pixel values ​​of the processed image used to create the machine learning model (i.e., the pixel values ​​of the band region and background region) are not particularly limited as long as they are capable of generating an inference image that can be binarized.

[0028] In order to facilitate the binarization process of the inference image, it is preferable that the processed image satisfies the following formulas (1) to (3).

[0029] P A ≧Pth2 …(1) P th1 <P B <P th2 …(2) P C ≦P th1 …(3) however, P A is the pixel value of the pixel A, P B is the pixel value of the pixel B other than the pixel A, P C is the pixel value of pixel C corresponding to the background region, P th1 is the first threshold, P th2 is the second threshold.

[0030] In formulas (1) to (3), P th1 and P th2 The value of is not particularly limited, and an optimal value can be selected depending on the purpose. P th1 can be zero or greater than zero. th2 is the maximum value of the gradation of the processed image P max Alternatively, P max A smaller value is also acceptable. For example, P th1 P max 30% of P th2 P max It may be 70% of that.

[0031] The band region is, in particular, pixel A out From pixel C ad It is preferable that the pixel values ​​decrease linearly toward the pixel value . In addition, for all grains included in the processed image, the pixel value is P max (Pixel A out ) to zero (pixel C ad It is preferable that the image density is linearly decreased to 1 / 1000. A method for generating a processed image including such a band region will be described later.

[0032] [1.2.3. Pixel value gradient] [A. Overview] The "pixel value gradient" refers to the ratio (|Δy| / Δx) of the absolute value (|Δy|) of the difference in pixel values ​​between two pixels spaced apart to the distance (Δx) between the two pixels. "Distance between two pixels" refers to the center-to-center distance between the pixels. The "pixel value gradient of the band region" is pixel A out and pixel B out Pixel C is located adjacent to the outside of ad The minimum distance between (Δx min ) for pixel A out Pixel value of pixel C ad The absolute value of the difference between the pixel value of (|Δy AC |) ratio (|y AC | / Δx min )

[0033] In the processed image, pixels A are arranged linearly inside the band region, and there is a pixel value gradient on both sides of the linearly arranged pixels A. The pixel value gradient of the band region is correlated with the width of the band region, and the larger the pixel value gradient of the band region, the narrower the width of the band region. Therefore, if the pixel value gradient of the band region included in the processed image is excessively large, the width of the band region included in the inferred image is excessively narrow. When such an inferred image is subjected to binarization and thinning processing, some of the lines corresponding to the grain boundaries may be missing in the grain boundary image.

[0034] Furthermore, when the pixel value gradients on both sides of the grain boundary are the same, pixel A is positioned approximately on the center line of the band region. When an inference image is created using such a processed image and then the inference image is subjected to binarization and thinning processes, the lines representing the grain boundaries appear in the grain boundary image at positions approximately equal to the positions of the original grain boundaries.

[0035] On the other hand, if the pixel value gradients on both sides of the grain boundary are not the same, pixel A is positioned away from the center line of the band region to the side with the larger pixel value gradient. When an inference image is created using such a processed image and then subjected to binarization and thinning processes, the center line of the band region is extracted as the "line representing the grain boundary." That is, if the pixel value gradient on both sides of the grain boundary is not the same, the line representing the grain boundary in the grain boundary image is likely to be biased from the original position of the grain boundary to the side with the smaller pixel value gradient.

[0036] Therefore, if the pixel value gradient of the band region included in the processed image becomes excessively small, when the inference image is binarized and thinned, the difference between the position of the line representing the grain boundary and the position of the original grain boundary may become large. In particular, when the grain size varies greatly, grains with large grain sizes may be displayed smaller than their actual size in the grain boundary image. It is preferable to select an optimal value for the pixel value gradient of the band region taking these points into consideration.

[0037] [B. Example of pixel value gradient in band region (1)] The processed images are (a) The pixel value of pixel A is greater than the pixel value of pixel C, and (b) The pixel value gradient of the band region is equal to the pixel value gradient obtained by binarizing and distance transforming the tissue image. But that's okay.

[0038] In general, a "distance transform" refers to a transform that gives the shortest distance from an arbitrary point to the nearest pixel with a pixel value of zero. However, in the present invention, a "distance transform" refers to a transform that gives the shortest distance from an arbitrary point to the nearest specific pixel (i.e., pixel A out ) is a transformation that gives the shortest distance to In the present invention, the method of distance transformation is not particularly limited. Examples of the distance transformation method include Euclidean distance transformation, pseudo-Euclidean distance transformation, 4-neighbor distance transformation, 8-neighbor distance transformation, etc. As the distance transformation, the Euclidean distance transformation is particularly preferable.

[0039] Even when a texture image is used as input data and a distance-transformed image is used as ground truth data, a machine learning model can be constructed. By inputting a texture image to such a machine learning model, an inferred image (distance-transformed image) can be output. Furthermore, by appropriately setting the threshold value during binarization processing, the inferred image (distance-transformed image) can also be converted into a grain boundary image.

[0040] The distance transformation image is pixel A out and pixel C ad In order to improve the accuracy of grain boundary discrimination, it is preferable to make the pixel value gradient of the band region in the processed image larger than that in the distance transformed image.

[0041] [C. Example of pixel value gradient in band region (2)] In the processed image, it is preferable that the pixel value gradient of the band region within each crystal grain is equal to or greater than the pixel value gradient based on the maximum crystal grain and equal to or less than the pixel value gradient based on the minimum crystal grain.

[0042] "Pixel value gradient based on the largest crystal grain" refers to the pixel value gradient of the band region at the time when pixel C corresponding to at least one background region first appears within the crystal grain of the largest grain size, when the pixel value gradient of the band region within the crystal grain of the largest grain size contained in the tissue image is gradually increased so that the pixel value gradient of the band region within the crystal grain of the largest grain size becomes uniform. "Pixel value gradient based on the smallest grain" refers to the pixel value gradient of the band region at the time when at least one pixel C first appears within the smallest grain, when the pixel value gradient of the band region within the smallest grain included in the structure image is gradually increased so that the pixel value gradient of the band region within the smallest grain becomes uniform.

[0043] As described later, for tissue images, (a) A method of binarizing a tissue image so that pixels corresponding to grain boundaries are white and pixels corresponding to background regions are black, and then a method of increasing the pixel value gradient of the resulting binarized image. (b) A case in which the texture image is binarized so that pixels corresponding to grain boundaries are black and pixels corresponding to background regions are white, and the resulting binarized image is then subjected to a process of increasing the pixel value gradient. There is. In the former case, if the pixel value in the grain is greater than the first threshold P th1 The process of increasing the pixel value gradient is performed until a pixel C appears such that On the other hand, in the latter case, the pixel value is the maximum value P max -First threshold P th1 The process of increasing the pixel value gradient is carried out until a pixel C appears such that "

[0044] The pixel value gradients of the band regions within each crystal grain contained in the processed image may be the same as or different from each other, as long as they are equal to or greater than the pixel value gradient based on the maximum crystal grain and equal to or less than the pixel value gradient based on the minimum crystal grain. In addition, the pixel value of pixel A included in each band region (i.e., the second threshold P th2 ) is P as long as the above conditions are met. max Alternatively, P max In order to improve the accuracy of grain boundary discrimination, the pixel value of pixel A for all crystal grains is P max is preferred.

[0045] Similarly, pixel B in each band region out The pixel values ​​of,P,are,P,, respectively, as long as the inference image can be binarized. th1 Alternatively, the pixel value may be one gray level higher than P th1 A pixel value that is two or more gradations larger may also be used. Furthermore, the pixel value of pixel C in each crystal grain may be zero or may be a pixel value greater than zero, so long as the above condition is satisfied.

[0046] As described above, when an inference image is created using a processed image with an excessively small pixel value gradient in the band region, and the inference image is then binarized and thinned, the difference between the position of the line representing the grain boundary and the position of the original grain boundary may become large in the grain boundary image. Therefore, it is preferable that the pixel value gradient of the band region in each crystal grain is equal to or greater than the pixel value gradient of the maximum crystal grain standard. On the other hand, when an inference image is created using a processed image in which the pixel value gradient of the band region is excessively large, and the inference image is then binarized and thinned, a part of the line representing the grain boundary may be missing in the grain boundary image. Therefore, it is preferable that the pixel value gradient of the band region in each crystal grain is equal to or less than the pixel value gradient of the minimum crystal grain standard.

[0047] [D. Example of pixel value gradient (3)] The processed images are For each band region included in each crystal grain A whose grain size is in the range from the median grain size to the maximum grain size, the pixel value gradient of the band region is a pixel value gradient based on the median grain size, For the band region included in each crystal grain B whose grain size is in the range of the minimum grain size or more and less than the median grain size, the pixel value gradient of the band region is the pixel value gradient based on the crystal grain B. The one is preferred.

[0048] The term "median particle size" refers to a specific particle size in the range from the minimum particle size to the maximum particle size. The "minimum grain size" refers to the grain size of the smallest crystal grain among the crystal grains contained in the structure image. The "maximum grain size" refers to the grain size of the largest crystal grain among the crystal grains contained in the structure image. The "grain size" refers to the circle equivalent diameter.

[0049] The median particle size is not particularly limited, and an optimal value can be selected depending on the purpose. Specific examples of the median particle size include the following. Any of them may be used in the present invention. (a) Average particle size (=Σ particle size x frequency), (b) Most frequent particle size, (c) median particle size; (d) area-weighted mean particle size, (e) the particle size that provides the greatest degree of separation when dividing the particle size histogram in half; (f) Particle size that maximizes the degree of separation when dividing the particle size area ratio graph in half

[0050] "Pixel value gradient based on medium grain" refers to the pixel value gradient of the band region within a medium grain size crystal grain at the time when at least one pixel C corresponding to a background region first appears within the medium grain size crystal grain when the pixel value gradient of the band region within the medium grain size crystal grain contained in the tissue image is gradually increased so that the pixel value gradient of the band region within the medium grain size crystal grain is uniform. The "pixel value gradient based on crystal grain B" refers to the pixel value gradient of the band region at the time when at least one pixel C first appears within crystal grain B when the pixel value gradient of the band region within crystal grain B included in the tissue image is gradually increased so that the pixel value gradient of the band region within crystal grain B becomes uniform.

[0051] The details of the method for making at least one pixel C appear are as described above, and therefore will not be described here.

[0052] When the pixel value gradient of each band region contained in each crystal grain B is adjusted so that the pixel value gradient of the band region becomes the pixel value gradient based on the crystal grain B, the pixel value gradient of the band region of each crystal grain B has a different value depending on the grain size. The pixel value of pixel A included in each band region (i.e., P th2 ) is P max Alternatively, P max In addition, the pixel value of the pixel C in each crystal grain may be zero or may be a pixel value greater than zero.

[0053] When the variation in particle size is small, the pixel value of pixel A is P maxWhen the pixel value gradient of the band region in each crystal grain is adjusted so that an extremely small number of pixels with a pixel value of zero appear within the crystal grain, the pixel value gradient of the band region on both sides of the grain boundary becomes almost uniform. When an inferred image is output using a machine learning model constructed using such a processed image, and the inferred image is subjected to binarization and thinning processing, the line representing the grain boundary is unlikely to deviate from the original position of the grain boundary.

[0054] On the other hand, when the particle size variation is large, the pixel value of pixel A is P max When the pixel value gradient of the band region in each crystal grain is adjusted so that an extremely small number of pixels with a pixel value of zero appear within the crystal grain, the pixel value gradient of the band region on both sides of the grain boundary often becomes uneven. In this case, when the inference image is subjected to binarization and thinning processing, the line representing the grain boundary of a crystal grain with a large grain size is likely to shift from the original position of the grain boundary. For example, when a large grain and a small grain are adjacent to each other, the line representing the grain boundary is shifted from its original position. As a result, the grain size of the large grain in the processed image is smaller than the grain size in the texture image. On the other hand, the grain size of the small grain in the processed image is larger than the grain size in the texture image.

[0055] In such a case, it is preferable to optimize the pixel value gradient of the band region according to the grain size, rather than adjusting the pixel value gradient of the band region on a uniform basis for all crystal grains. Specifically, for each crystal grain A whose grain size is equal to or larger than the median grain size, the pixel value gradient of the band region is made to coincide with the pixel value gradient based on the median grain size, regardless of the grain size. This makes it possible to suppress the misalignment of the lines representing the grain boundaries in the grain boundary image, even when the grain sizes vary widely. On the other hand, for each crystal grain B whose grain size is less than the median grain size, the pixel value gradient of the band region is changed according to the grain size. Specifically, the pixel value gradient based on the crystal grain B is selected as the pixel value gradient of the band region of each crystal grain B. This makes it possible to minimize the loss of part of the line representing the grain boundary in the grain boundary image.

[0056] [2. Dataset (2)] The data set according to the second embodiment of the present invention is Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; It is equipped with: In this embodiment, the pixel value of the band region is determined based on the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out is gradually increasing towards

[0057] [2.1. Tissue images] The details of the tissue image are similar to those in the first embodiment, and therefore will not be described.

[0058] [2.2. Processed images] [2.2.1. Definition] The term "processed image" refers to an image that includes a band region and a remaining background region. The processed image is obtained by performing a predetermined processing process on a tissue image. The processed image is used as the correct answer data when performing machine learning. In this embodiment, the "predetermined processing" refers to a process of adjusting the pixel values ​​and pixel value gradients of the tissue image so that grain boundaries are converted into band regions and the band regions are displayed darker than the background region.

[0059] The "band region" refers to a region in which the width of the grain boundary included in the structure image is expanded toward the inside of two adjacent crystal grains. Among the pixels included in the band region in this embodiment, the pixels with the smallest pixel value are all pixels A, and the pixels with the largest pixel value are all pixels B. out Furthermore, the pixel value of the band region is pixel A out From pixel B out is gradually increasing towards "Background region" refers to the region remaining after removing the band regions from the tissue image.

[0060] "Pixel value is gradually increasing" means that pixel A out From pixel B out The band area is a region where the pixel values ​​increase continuously or stepwise from pixel A to pixel B. out From pixel B out The pixel value may increase linearly or may increase in a curved manner toward the pixel value. Other points regarding the definition of the processed image are the same as those in the first embodiment, so the description will be omitted.

[0061] [2.2.2. Pixel value] After generating a machine learning model using the data set according to the present embodiment, when an arbitrary tissue image is input to the machine learning model, an inference image in which grain boundaries are converted into band regions is output. The inference image is further subjected to binarization and thinning processes to become a grain boundary image (binarized image). Therefore, the pixel values ​​of the processed image used to create the machine learning model (i.e., the pixel values ​​of the band region and background region) are not particularly limited as long as they are capable of generating an inference image that can be binarized.

[0062] In order to facilitate the binarization process of the inference image, it is preferable that the processed image satisfies the following formulas (4) to (6).

[0063] P C ≧P th4 …(4) P th3 <P B <P th4 …(5) P A ≦P th3 …(6) however, P A is the pixel value of the pixel A, P B is the pixel value of the pixel B other than the pixel A, P C is the pixel value of pixel C corresponding to the background region, P th3 is the third threshold, P th4is the fourth threshold.

[0064] In formulas (4) to (6), P th3 and P th4 The value of is not particularly limited, and an optimal value can be selected depending on the purpose. P th3 can be zero or greater than zero. th4 is the maximum value of the gradation of the processed image P max Alternatively, P max A smaller value is also acceptable. For example, P th3 P max 30% of P th4 P max It may be 70% of that.

[0065] The band region is, in particular, pixel A out From pixel C ad It is preferable that the pixel values ​​increase linearly toward the pixel value . In addition, the pixel value of all grains in the processed image is zero (pixel A out ) to P max (Pixel C ad ) is preferably increased linearly. A method for generating a processed image including such a band region will be described later.

[0066] [2.2.3. Pixel value gradient] [A. Overview] The details of the pixel value gradient are the same as those in the first embodiment, and therefore will not be described here.

[0067] [B. Example of pixel value gradient in band region (4)] The processed images are (a) The pixel value of pixel A is smaller than the pixel value of pixel C, and (b) The pixel value gradient of the band region is equal to the pixel value gradient obtained by binarizing and distance transforming the tissue image. But that's okay.

[0068] Specific example (4) corresponds to specific example (1) that has been subjected to black and white inversion processing. That is, in the processed image of specific example (4), the grain boundaries are displayed in black and the grain interiors are displayed in white. Other points regarding specific example (4) are the same as those of specific example (1), so the explanation will be omitted.

[0069] [C. Example of pixel value gradient in band region (5)] In the processed image, it is preferable that the pixel value gradient of the band region within each crystal grain is equal to or greater than the pixel value gradient based on the maximum crystal grain and equal to or less than the pixel value gradient based on the minimum crystal grain.

[0070] "Pixel value gradient based on the largest crystal grain" refers to the pixel value gradient of the band region at the time when pixel C corresponding to at least one background region first appears within the crystal grain of the largest grain size, when the pixel value gradient of the band region within the crystal grain of the largest grain size contained in the tissue image is gradually increased so that the pixel value gradient of the band region within the crystal grain of the largest grain size becomes uniform. "Pixel value gradient based on the smallest grain" refers to the pixel value gradient of the band region at the time when at least one pixel C first appears within the smallest grain, when the pixel value gradient of the band region within the smallest grain included in the structure image is gradually increased so that the pixel value gradient of the band region within the smallest grain becomes uniform.

[0071] As described later, for tissue images, (a) A method of binarizing a tissue image so that pixels corresponding to grain boundaries are white and pixels corresponding to background regions are black, and then a method of increasing the pixel value gradient of the resulting binarized image. (b) A case in which the texture image is binarized so that pixels corresponding to grain boundaries are black and pixels corresponding to background regions are white, and the resulting binarized image is then subjected to a process of increasing the pixel value gradient. There is. In the former case, the pixel value in the grain is the maximum value P max - 4th threshold P th4 The process of increasing the pixel value gradient is carried out until a pixel C appears such that " On the other hand, in the latter case, there is a pixel value Pth4 The process of increasing the pixel value gradient is performed until a pixel C appears such that

[0072] The pixel value gradients of the band regions within each crystal grain contained in the processed image may be the same as or different from each other, as long as they are equal to or greater than the pixel value gradient based on the maximum crystal grain and equal to or less than the pixel value gradient based on the minimum crystal grain. In addition, the pixel value of pixel A included in each band region (i.e., P th3 ) may be zero or may be a value greater than zero as long as the above condition is satisfied. In order to improve the accuracy of grain boundary discrimination, it is preferable that the pixel value of pixel A is zero for all crystal grains.

[0073] Similarly, pixel B in each band region out The pixel values ​​of,P,are,P,, respectively, as long as the inference image can be binarized. th4 Alternatively, the pixel value may be one gray level lower than P th4 A pixel value that is two or more gradations smaller may also be used. Furthermore, the pixel value of pixel C in each grain is, as long as the above condition is satisfied, P max Or, P max A smaller pixel value may also be used.

[0074] Specific example (5) corresponds to specific example (2) that has been subjected to black and white inversion processing. That is, in the processed image of specific example (5), the grain boundaries are displayed in black and the grain interiors are displayed in white. Other points regarding specific example (5) are similar to specific example (2), so the explanation will be omitted.

[0075] [D. Example of pixel value gradient (6)] The processed images are For each band region included in each crystal grain A whose grain size is in the range from the median grain size to the maximum grain size, the pixel value gradient of the band region is a pixel value gradient based on the median grain size, For the band region included in each crystal grain B whose grain size is in the range of the minimum grain size or more and less than the median grain size, the pixel value gradient of the band region is the pixel value gradient based on the crystal grain B. The one is preferred.

[0076] "Pixel value gradient based on medium grain" refers to the pixel value gradient of a band region within a medium grain size crystal grain contained in a tissue image at the time when pixel C corresponding to at least one background region first appears within the medium grain size crystal grain, when the pixel value gradient of the band region within the medium grain size crystal grain is gradually increased so that the pixel value gradient of the band region within the medium grain size crystal grain contained in the tissue image becomes uniform. The "pixel value gradient based on crystal grain B" refers to the pixel value gradient of the band region at the time when at least one pixel C first appears within crystal grain B when the pixel value gradient of the band region within crystal grain B included in the tissue image is gradually increased so that the pixel value gradient of the band region within crystal grain B becomes uniform.

[0077] The details of the method for making at least one pixel C appear are as described above, and therefore will not be described here.

[0078] When the pixel value gradient of each band region contained in each crystal grain B is adjusted so that the pixel value gradient of the band region becomes the pixel value gradient based on the crystal grain B, the pixel value gradient of the band region of each crystal grain B has a different value depending on the grain size. The pixel value of pixel A included in each band region (i.e., P th3 ) may be zero or may be a value greater than zero. The pixel value of pixel C in each grain is max Alternatively, P max A smaller pixel value may also be used.

[0079] Specific example (6) corresponds to specific example (3) that has been subjected to black and white inversion processing. That is, in the processed image of specific example (6), the grain boundaries are displayed in black and the grain interiors are displayed in white. Other points regarding specific example (6) are similar to specific example (3), so the explanation will be omitted.

[0080] [3. How to generate processed images] The processed image can be generated by various methods. In order to minimize the positional shift of the lines representing the grain boundaries and the partial loss of the lines representing the grain boundaries in the grain boundary image, the processed image is (a) In each grain, the pixel value of the band region is pixel A out From pixel B out Pixel C adjacent to ad and (b) In each grain, the pixel value of pixel A is P max and the pixel value of pixel C is zero. is preferred.

[0081] The processed image including the band region that satisfies these conditions is (A) The structure image is binarized so that the pixel values ​​of pixels corresponding to grain boundaries are the maximum value of the gradation (white) and the pixel values ​​of pixels corresponding to inside the crystal grains are zero (black) (Step A). (B) The obtained binarized image is subjected to black and white inversion processing (step B). (C) A distance transform is performed on the obtained black and white inverted image (step C). (D) performing a process for increasing the pixel value gradient of the band region for each crystal grain so as to obtain a band region having a predetermined width on the obtained distance transformation image (step D); (E) The obtained gradient augmented image is subjected to a black and white conversion process (step E). This can be obtained by the following method (first method).

[0082] Alternatively, the processed image may be (A') The structure image is binarized so that the pixel values ​​of pixels corresponding to grain boundaries are zero (black) and the pixel values ​​of pixels corresponding to inside crystal grains are the maximum gradation value (white) (step A'); (C') performing distance transformation on the obtained binarized image (step C'); (D) performing a process for increasing the pixel value gradient of the band region for each crystal grain so as to obtain a band region having a predetermined width on the obtained distance transformation image (step D); (E) The obtained gradient augmented image is subjected to a black and white conversion process (step E). This can be obtained by the following method (second method).

[0083] In the first and second methods, the order of steps D and E may be interchanged. In addition, in the first and second methods, if step E is omitted, out From C ad The resulting processed image has pixel values ​​that increase linearly from

[0084] FIG. 1 shows a schematic diagram of a method for generating a processed image in which the grain boundaries are converted from a structural image containing grain boundaries into a band-like region that is wider than the grain boundaries. The upper left diagram in FIG. 1 is an image that has been binarized (step A) so that pixels corresponding to grain boundaries are white and pixels corresponding to inside crystal grains are black. The method of binarization is not particularly limited, and any known method can be used. When binarization is performed using an image analyzer, a portion of the line representing the grain boundary is often missing. Therefore, when binarization is performed using an image analyzer, manual marking of the binarized image is required.

[0085] The upper right diagram in FIG. 1 is an image obtained by inverting the binarized image (step B), performing distance transformation on the binarized image (step C), and further inverting the distance-transformed image (step E). As described above, the "distance transformation" in the present invention refers to the process of finding the nearest specific pixel (i.e., pixel A) from an arbitrary point. out ) is a transformation that gives the shortest distance to the pixel A. A distance function is used to calculate the distance. The method of distance transformation is not particularly limited, and any known method can be used. In the upper right diagram of FIG. 1, pixel A out The further away a pixel is, the darker it appears.

[0086] Next, a process for increasing the pixel value gradient of the band region is performed on the distance transformed image (step D). The method for increasing the pixel value gradient of the band region is not particularly limited. As a method for increasing the pixel value gradient of the band region, for example, (a) A method of increasing the pixel value gradient of the band region so that the pixel value gradient of the band region for all grains matches the pixel value gradient of the maximum grain reference (lower left diagram in Figure 1), (b) A method of increasing the pixel value gradient of the band region so that the pixel value gradient of the band region for all grains matches the pixel value gradient of the minimum grain standard (bottom center diagram in Figure 1), (c) A method of increasing the pixel value gradient of the band region so that the pixel value gradient of the band region for all grains matches the pixel value gradient of the medium grain reference (bottom right diagram in Figure 1), etc.

[0087] In addition, when the method of making the pixel value gradient of the band region coincide with the pixel value gradient of the maximum grain standard for all grains (lower left diagram in Figure 1) is used, the minimum pixel value is not zero in grains other than the maximum grain. In this case, it is preferable to further increase the pixel value gradient of the band region for each grain other than the maximum grain until the minimum pixel value in the grain becomes zero. Setting the minimum pixel value in the grain to zero makes it easier to binarize the inference image.

[0088] This is also the case when the method of making the pixel value gradient of the band region coincide with the pixel value gradient of the medium crystal grain standard for all crystal grains (lower right diagram in Fig. 1) is used. That is, for each crystal grain B having a grain size smaller than the medium grain size, it is preferable to increase the pixel value gradient of the band region of each crystal grain B so that the pixel value gradient of the band region coincides with the pixel value gradient of the crystal grain B standard.

[0089] Fig. 2 shows an example of a tissue image and a binary image. The left diagram in Fig. 2 is a tissue image. The upper right diagram in Fig. 2 is a binary image obtained by binarizing the tissue image. The lower right diagram in Fig. 2 is a schematic diagram of changes in pixel values ​​in the binary image.

[0090] FIG. 3 shows an example of a distance transformed image, a gradient augmented image, and a gradient adjusted image. The upper left diagram in Fig. 3 is a distance-transformed image obtained by performing black-and-white conversion → distance conversion → black-and-white conversion on the binarized image in the right diagram in Fig. 2. The lower left diagram in Fig. 3 is a schematic diagram of the change in pixel values ​​in the distance-transformed image. The upper middle diagram in Fig. 3 is a gradient augmented image obtained by performing a process to increase the pixel value gradient of the band region in the distance transformation image so that the pixel value gradient of the band region in all grains matches the pixel value gradient of the medium grain reference. The lower middle diagram in Fig. 3 is a schematic diagram of the change in pixel value in the gradient augmented image. The upper right diagram in Fig. 3 is a gradient-adjusted image obtained by performing a process on the gradient-increased image to further increase the pixel value gradient of the band region only for crystal grains whose minimum pixel value is not zero. The lower right diagram in Fig. 3 is a schematic diagram of the change in pixel value in the gradient-adjusted image.

[0091] [4. Machine Learning Model] The machine learning model according to the present invention is obtained by training an artificial intelligence to learn machine learning from the dataset according to the present invention. The details of the dataset are as described above, and therefore will not be described here. A known method can be used to train an artificial intelligence to learn the dataset by machine learning.

[0092] [5. Tissue image analysis program] FIG. 4 shows a flow diagram of the tissue image analysis program according to the present invention. First, in step 1 (hereinafter simply referred to as "S1"), a structure image including crystal grains and grain boundaries is stored in a memory (procedure A). Next, in S2, the tissue image is input to the machine learning model according to the present invention that is stored in advance in the memory, and an inferred image including the band region and the background region is output from the machine learning model, and the inferred image is stored in the memory (step B). Details of the machine learning model and the data set for generating the model are as described above, and therefore will not be described here.

[0093] Next, in S3, a binarization process is performed on the inference image to obtain a binarized image including a binarized band region, and the binarized image is stored in a memory (step C). The method of the binarization process is not particularly limited, and any known method can be used. The threshold value for the binarization process is not particularly limited, and an optimal value can be selected depending on the purpose. For example, if the pixel value of pixel C corresponding to the background region is zero, the threshold value may be zero or may be a value greater than zero.

[0094] Next, in S4, a grain boundary image in which the band region is converted into a line is obtained by performing a line thinning process on the binary image, and the grain boundary image is stored in the memory (step D). "Line thinning process" refers to a process of extracting a figure corresponding to the center line from a figure with width. The method of line thinning process is not particularly limited, and a known method can be used.

[0095] Next, the process proceeds to S5. In S5, it is determined whether or not to perform whisker removal processing. When a band region is thinned, unintended short lines called "whiskers" may be output due to shape noise present at the edge of the band region. If whiskers are included in the line obtained by thinning the band region, and there is a risk that the presence of the whiskers may adversely affect the accuracy of grain boundary discrimination (S5: YES), the process proceeds to S6.

[0096] In S6, a whisker removal process is performed on the grain boundary image to obtain a corrected grain boundary image in which whiskers are removed from the lines, and the corrected grain boundary image is stored in memory (step E). The method of the whisker removal process is not particularly limited, and a known method can be used. Then, the process proceeds to S7. On the other hand, if the line obtained by thinning the band region does not include whiskers, or if there is no risk that the presence of whiskers will adversely affect the accuracy of grain boundary discrimination (S5: NO), the process proceeds directly to S7.

[0097] In S7, it is determined whether or not to continue the process. If the process is to be continued (S7: YES), the process returns to S1. Then, the above-mentioned steps S1 to S7 are repeated. On the other hand, if the process is not to be continued (S7: NO), the process is terminated. It should be noted that the whisker removal process is not necessarily required. If the whisker removal process is not performed, steps S5 and S6 can be omitted.

[0098] FIG. 5 shows an example of an inferred image, a binarized image, and a corrected grain boundary image created using the program for analyzing texture images according to the present invention. The upper left diagram in Fig. 5 is an inferred image obtained by inputting a structural image into the machine learning model according to the present invention. The upper right diagram in Fig. 5 is a binarized image obtained by performing binarization processing on the inferred image. The lower diagram in Fig. 5 is a corrected grain boundary image obtained by performing thinning processing and whisker removal processing on the binarized image.

[0099] [5. Effect] A machine learning model is obtained by machine learning a structure image including grain boundaries as input data and a grain boundary image obtained by binarizing the structure image as correct answer data. When an arbitrary structure image is input to the obtained machine learning model, a grain boundary image corresponding to the input structure image is output. However, the grain boundary image obtained in this way may be missing some of the lines representing the grain boundaries.

[0100] In contrast, a machine learning model can be obtained by using a structural image including grain boundaries as input data and a processed image including a band region in which the grain boundaries have been widened as correct answer data for machine learning. When an arbitrary structural image is input to the obtained machine learning model, an inferred image in which the linear grain boundaries have been converted into a wide band region is output. The obtained inferred image is subjected to a binarization process, and the binarized band region is subjected to a thinning process, resulting in a grain boundary image corresponding to the input structural image. The grain boundary image obtained in this way is less likely to lose parts of the lines representing the grain boundaries than grain boundary images obtained by the conventional method, which is believed to be because the grain boundaries are converted into wide band regions before being binarized.

[0101] Furthermore, if the pixel value gradient of the band region (i.e., the width of the band region) is inappropriate, the position of the line may be significantly shifted from the original position of the grain boundary when the band region is converted back into a line. The magnitude of the shift increases as the variation in grain size of each crystal grain in the structure image increases. As a result, when the variation in grain size of the crystal grains is large, a crystal grain with a large grain size may be displayed in the grain boundary image as a crystal grain smaller than its actual size.

[0102] In contrast, optimizing the pixel value gradient of the band region can reduce the difference between the position of the line on the grain boundary image and the position of the original grain boundary. This can prevent large grains from being displayed smaller than their actual size in the grain boundary image. This can also prevent a decrease in the evaluation accuracy of the number of grains, grain size, grain size distribution, etc. EXAMPLES

[0103] Comparative Example 1 [1. Test method] Tissue images were used as input data. The correct answer data was a binarized image obtained by binarizing the structure image as it was. The binarized image of Comparative Example 1 was an image in which grain boundaries were displayed as narrow white lines, that is, a grain boundary image. When the structure image was binarized using an image analyzer, some of the lines representing the grain boundaries were missing, so the binarized image was corrected by manual marking.

[0104] Using the multiple data sets thus obtained, artificial intelligence was trained to perform machine learning to obtain a machine learning model. Next, one of the structural images used as input data was input to the obtained machine learning model, and an inference image was output. In the case of Comparative Example 1, the inference image was an image in which grain boundaries were displayed as narrow white lines, i.e., a grain boundary image.

[0105] [2. Results] Fig. 6 shows an example of the data set used in Comparative Example 1, and an example of an inferred image created based on the machine learning model of Comparative Example 1. The upper left diagram in Fig. 6 is the input data (structural image). The upper right diagram in Fig. 6 is the correct answer data (grain boundary image). The lower diagram in Fig. 6 is an inferred image (grain boundary image) obtained by inputting the structural image in the upper left diagram in Fig. 6 into the machine learning model. We compared the input data (structure image) in the upper left diagram of Fig. 6 with the inferred image (grain boundary image) in the lower diagram of Fig. 6. As a result, we found that the lines representing the grain boundaries were partially missing.

[0106] Example 1 [1. Test method] Tissue images were used as input data. The correct answer data is (a) The structure image is binarized so that grain boundaries are white (i.e., pixel values ​​are the maximum value of the gray scale) and background regions are black (i.e., pixel values ​​are zero); (b) The binarized image is inverted. (c) The black and white inverted image is subjected to distance transformation. (d1) increasing the pixel value gradient so that the pixel value gradient of the band region of each crystal grain becomes the pixel value gradient of the maximum crystal grain reference; (d2) For each crystal grain B other than the largest crystal grain, the pixel value gradient is adjusted so that the pixel value gradient of the band region becomes the pixel value gradient based on the crystal grain B; (e) The gradient-adjusted image is then inverted. The material obtained by this method was used.

[0107] Using the multiple data sets thus obtained, artificial intelligence was trained to perform machine learning, and a machine learning model was obtained. Next, one of the structure images used as input data was input to the obtained machine learning model, and an inference image was output. In the case of Example 1, the inference image is an image in which the grain boundary is converted into a wide band region. Furthermore, the obtained inference image was subjected to binarization processing and thinning processing to obtain a grain boundary image.

[0108] [2. Results] FIG. 7 shows an example of the data set used in the first embodiment, as well as examples of an inferred image, a binarized image, and a grain boundary image created based on the machine learning model of the first embodiment. The upper left image in Fig. 7 shows the input data (structural image). The upper right image in Fig. 7 shows the correct data (processed image) obtained by converting the grain boundaries in the structural image into a wide band region. The lower left diagram in Fig. 7 is an inferred image obtained by inputting the structure image in the upper left diagram in Fig. 7 into a machine learning model. The lower center diagram in Fig. 7 is a binary image obtained by binarizing the inferred image. Furthermore, the lower right diagram in Fig. 7 is a grain boundary image obtained by performing thinning processing on the binary image.

[0109] The input data (structure image) in the upper left diagram of Fig. 7 was compared with the grain boundary image in the lower right diagram of Fig. 7. As a result, almost no missing lines representing grain boundaries were observed in the grain boundary image. However, in the grain boundary images, the large grains were displayed smaller than their actual size, which is believed to be because the pixel value gradient of the band region within the large grains is smaller than that of the adjacent grains.

[0110] Example 2 [1. Test method] Tissue images were used as input data. The correct answer data is (a) The structure image is binarized so that grain boundaries are white (i.e., pixel values ​​are the maximum value of the gray scale) and background regions are black (i.e., pixel values ​​are zero); (b) The binarized image is inverted. (c) The black and white inverted image is subjected to distance transformation. (d1) for each band region included in each crystal grain A whose grain size is in the range of the median grain size or more and the maximum grain size or less, the pixel value gradient is increased so that the pixel value gradient of the band region coincides with the pixel value gradient of the median grain size standard; (d2) for each crystal grain B having a grain size in the range of not less than the minimum grain size and not more than the median grain size, the pixel value gradient is adjusted so that the pixel value gradient of the band region becomes the pixel value gradient based on the crystal grain B; (e) Further invert the gradient adjusted image The material obtained by this method was used.

[0111] Using the multiple data sets thus obtained, artificial intelligence was trained to perform machine learning, and a machine learning model was obtained. Next, one of the structure images used as input data was input to the obtained machine learning model, and an inference image was output. In the case of Example 2, the inference image is an image in which the grain boundary is converted into a wide band region. Furthermore, the obtained inference image was subjected to binarization processing and thinning processing to obtain a grain boundary image.

[0112] [2. Results] FIG. 8 shows an example of a data set used in the second embodiment, and examples of an inferred image, a binarized image, and a grain boundary image created based on the machine learning model of the second embodiment. The upper left image in Fig. 8 shows the input data (structural image). The upper right image in Fig. 8 shows the correct data (processed image) obtained by converting the grain boundaries in the structural image into a wide band region. The lower left diagram in Fig. 8 is an inferred image obtained by inputting the structure image in the upper left diagram in Fig. 8 into a machine learning model. The lower center diagram in Fig. 8 is a binary image obtained by binarizing the inferred image. Furthermore, the lower right diagram in Fig. 8 is a grain boundary image obtained by performing thinning processing on the binary image.

[0113] The input data (structure image) in the upper left diagram of Fig. 8 was compared with the grain boundary image in the lower right diagram of Fig. 8. As a result, almost no missing lines representing grain boundaries were observed in the grain boundary image. In the case of Example 2, the large crystal grains were displayed almost the same size as their actual size in the grain boundary image. This is believed to be because the pixel value gradient of the medium crystal grain standard was used as the pixel value gradient of the band region of each crystal grain A with a grain size equal to or larger than the medium grain size, regardless of grain size.

[0114] Although the embodiment of the present invention has been described in detail above, the present invention is not limited to the above embodiment, and various modifications are possible without departing from the gist of the present invention. [Industrial Applicability]

[0115] The dataset of the present invention can be used to create a machine learning model for converting linear grain boundaries in a tissue image into broad band regions. The machine learning model and the program for analyzing structural images according to the present invention can be used to evaluate the number, grain size, grain size distribution, etc. of crystal grains contained in a structural image of a polycrystalline material.

Claims

1. Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; Equipped with The pixel value of the band region is the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out is gradually decreasing towards Dataset.

2. The data set according to claim 1 , wherein the processed image satisfies the following formulas (1) to (3). P A ≧P th2 …(1) P th1 <P B <P th2 …(2) P C ≦P th1 …(3) however, P A is the pixel value of the pixel A, P B is the pixel value of the pixel B other than the pixel A, P C is the pixel value of pixel C corresponding to the background region, P th1 is the first threshold, P th2 is the second threshold.

3. The data set according to claim 1 , wherein the processed image has a pixel value gradient of the band region within each of the crystal grains that is greater than or equal to a pixel value gradient based on a maximum crystal grain and less than or equal to a pixel value gradient based on a minimum crystal grain. however, The "pixel value gradient" refers to the ratio (|Δy| / Δx) of the absolute value (|Δy|) of the difference in pixel values ​​between two pixels spaced apart from each other to the distance (Δx) between the two pixels; The "pixel value gradient of the band region" refers to the pixel A out and the pixel B out Pixel C arranged adjacent to the outside of ad The minimum distance between (Δx min ) for the pixel A out and the pixel value of pixel C ad The absolute value of the difference between the pixel value of AC |) ratio (|y AC | / Δx min ) The "pixel value gradient based on the largest crystal grain" refers to a pixel value gradient of the band region at a time when at least one pixel C corresponding to the background region first appears in the crystal grain of the largest grain size in a case where the pixel value gradient of the band region in the crystal grain of the largest grain size included in the structure image is increased stepwise so that the pixel value gradient of the band region in the crystal grain of the largest grain size included in the structure image becomes uniform, The "pixel value gradient based on the smallest crystal grain" refers to the pixel value gradient of the band region at the time when at least one of the pixels C first appears within the crystal grain of the smallest grain size when the pixel value gradient of the band region within the crystal grain of the smallest grain size included in the structure image is gradually increased so that the pixel value gradient of the band region within the crystal grain of the smallest grain size becomes uniform.

4. The processed image is For each of the band regions included in each crystal grain A having a grain size in the range of not less than the median grain size but not more than the maximum grain size, the pixel value gradient of the band region is a pixel value gradient based on the median grain size, For the band region included in each crystal grain B whose grain size is in the range of the minimum grain size or more and less than the median grain size, the pixel value gradient of the band region is the pixel value gradient based on the crystal grain B. The data set of claim 1. however, The "median particle size" refers to a specific particle size that is greater than the minimum particle size and less than the maximum particle size, The "pixel value gradient" refers to the ratio (|Δy| / Δx) of the absolute value (|Δy|) of the difference in pixel values ​​between two pixels spaced apart from each other to the distance (Δx) between the two pixels; The "pixel value gradient of the band region" refers to the pixel A out and the pixel B out Pixel C arranged adjacent to the outside of ad The minimum distance between (Δx min ) for the pixel A out and the pixel value of pixel C ad The absolute value of the difference between the pixel value of AC |) ratio (|y AC | / Δx min ) The "pixel value gradient based on the medium grain" refers to the pixel value gradient of the band region at the time when at least one pixel C corresponding to the background region first appears in the medium grain when the pixel value gradient of the band region in the medium grain is increased stepwise so that the pixel value gradient of the band region in the medium grain included in the structure image becomes uniform, The "pixel value gradient based on crystal grain B" refers to the pixel value gradient of the band region at the time when at least one of the pixels C first appears within the crystal grain B when the pixel value gradient of the band region within the crystal grain B included in the tissue image is gradually increased so that the pixel value gradient of the band region within the crystal grain B becomes uniform.

5. Input data consisting of a structural image including crystal grains and grain boundaries; Correct answer data consisting of a processed image including a band region in which the grain boundary width is expanded toward the inside of two adjacent crystal grains and a remaining background region; Equipped with The pixel value of the band region is the pixel A located at the outermost edge of the pixel A corresponding to the grain boundary. out Pixel B located at the outermost edge of pixel B corresponding to the band region out is gradually increasing towards Dataset.

6. The data set according to claim 5 , wherein the processed image satisfies the following formulas (4) to (6): P C ≧P th4 …(4) P th3 <P B <P th4 …(5) P A ≦P th3 …(6) however, P A is the pixel value of the pixel A, P B is the pixel value of the pixel B other than the pixel A, P C is the pixel value of pixel C corresponding to the background region, P th3 is the third threshold, P th4 is the fourth threshold.

7. The data set according to claim 5, wherein the processed image has a pixel value gradient of the band region within each of the crystal grains that is greater than or equal to a pixel value gradient based on a maximum crystal grain and less than or equal to a pixel value gradient based on a minimum crystal grain. however, The "pixel value gradient" refers to the ratio (|Δy| / Δx) of the absolute value (|Δy|) of the difference in pixel values ​​between two pixels spaced apart from each other to the distance (Δx) between the two pixels; The "pixel value gradient of the band region" refers to the pixel A out and the pixel B out Pixel C arranged adjacent to the outside of ad The minimum distance between (Δx min ) for the pixel A out and the pixel value of pixel C ad The absolute value of the difference between the pixel value of AC |) ratio (|y AC | / Δx min ) The "pixel value gradient based on the largest crystal grain" refers to a pixel value gradient of the band region at a time when at least one pixel C corresponding to the background region first appears in the crystal grain of the largest grain size in a case where the pixel value gradient of the band region in the crystal grain of the largest grain size included in the structure image is increased stepwise so that the pixel value gradient of the band region in the crystal grain of the largest grain size included in the structure image becomes uniform, The "pixel value gradient based on the smallest crystal grain" refers to the pixel value gradient of the band region at the time when at least one of the pixels C first appears within the crystal grain of the smallest grain size when the pixel value gradient of the band region within the crystal grain of the smallest grain size included in the structure image is gradually increased so that the pixel value gradient of the band region within the crystal grain of the smallest grain size becomes uniform.

8. The processed image is For each of the band regions included in each crystal grain A having a grain size in the range of not less than the median grain size but not more than the maximum grain size, the pixel value gradient of the band region is a pixel value gradient based on the median grain size, For the band region included in each crystal grain B whose grain size is in the range of the minimum grain size or more and less than the median grain size, the pixel value gradient of the band region is the pixel value gradient based on the crystal grain B. The data set according to claim 5. however, The "median particle size" refers to a specific particle size that is greater than the minimum particle size and less than the maximum particle size, The "pixel value gradient" refers to the ratio (|Δy| / Δx) of the absolute value (|Δy|) of the difference in pixel values ​​between two pixels spaced apart from each other to the distance (Δx) between the two pixels; The "pixel value gradient of the band region" refers to the pixel A out and the pixel B out Pixel C arranged adjacent to the outside of ad The minimum distance between (Δx min ) for the pixel A out and the pixel value of pixel C ad The absolute value of the difference between the pixel value of AC |) ratio (|y AC | / Δx min ) The "pixel value gradient based on the medium grain" refers to the pixel value gradient of the band region at the time when at least one pixel C corresponding to the background region first appears in the medium grain when the pixel value gradient of the band region in the medium grain is increased stepwise so that the pixel value gradient of the band region in the medium grain included in the structure image becomes uniform, The "pixel value gradient based on crystal grain B" refers to the pixel value gradient of the band region at the time when at least one of the pixels C first appears within the crystal grain B when the pixel value gradient of the band region within the crystal grain B included in the tissue image is gradually increased so that the pixel value gradient of the band region within the crystal grain B becomes uniform.

9. The median particle size is (a) average particle size; (b) most frequent particle size; (c) median particle size; (d) area-weighted average particle size; (e) the particle size that provides the greatest degree of separation when dividing the particle size histogram in half; or (f) Particle size that provides the maximum degree of separation when dividing the particle size area ratio graph in half The data set according to claim 4 or 8,

10. A machine learning model obtained by training an artificial intelligence to machine-learn the dataset according to claim 1 or 5.

11. A program for tissue image analysis that causes a computer to carry out the following steps: (A) Step A for storing a structural image including crystal grains and grain boundaries in a memory. (B) Step B of inputting the tissue image into the machine learning model described in claim 10 that has been pre-stored in the memory, outputting an inferred image including a band region and a background region from the machine learning model, and storing the inferred image in the memory. (C) A step C of performing a binarization process on the inference image to obtain a binary image including the binarized band region, and storing the binary image in the memory. (D) A step D of performing a thinning process on the binary image to obtain a grain boundary image in which the band regions are converted into lines, and storing the grain boundary image in the memory.

12. The procedure includes: (E) a step of performing a whisker removal process on the grain boundary image when the line includes whiskers, thereby obtaining a corrected grain boundary image in which the whiskers are removed from the line, and storing the corrected grain boundary image in the memory; The tissue image analysis program according to claim 11, further comprising:

Images