Material structure identification methods

Abstract

To provide a material structure identification method capable of highly accurately identifying a plurality of types of material structures by luminance in a structure observation image including the plurality of types of material structures.SOLUTION: A material structure identification method includes in a structure observation image including a plurality of types of material structures and displayed in gradation: (1) a representative value extraction step of acquiring a mode and a half width of a raw histogram with respect to a luminance histogram of the structure observation image; (2) a normal distribution creation step of creating a normal distribution having the mode and the half width; (3) a difference extraction process of extracting an object obtained from subtracting the normal distribution from the raw histogram as a difference distribution; (4) a threshold value determination step of determining at least one of luminance values corresponding to an intersection of the normal distribution and difference distribution as a threshold value; and (5) a structure separation step of classifying the material structure in the raw histogram, by regarding, with a threshold as a boundary, a region on a high luminance side and a region on a low luminance side, as contributions of mutually different material structures.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to a method for identifying material structures, and more particularly, to a method for identifying material structures in an observation image of a structure including a plurality of material structures.

Background Art

[0002] In an observation image of a structure including a plurality of types of material structures, it is important to identify and detect each material structure in order to know the state of the material. For example, in a metal material, the distribution of material structures such as crystal phases has a great influence on the material properties, and observation images obtained by an optical microscope, an electron microscope, or the like are used to evaluate the structure. When a plurality of types of material structures, such as a plurality of crystal phases, are observed in those observation images, as a method for identifying a plurality of types of material structures on an image, a method for identifying a material structure by luminance in a grayscale image or the like is widely used. Typically, binarization is used as a method for identifying two types of material structures. As a method for determining a threshold value for binarizing a grayscale image, a plurality of methods have been proposed, such as those disclosed in Patent Documents 1 to 3.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0004] In microstructure observation images containing multiple material structures, the accuracy of material structure identification is greatly affected by the brightness threshold set for distinguishing different structures, such as through binarization. Setting an appropriate threshold is crucial for highly accurate material structure identification. However, when the brightness difference between multiple different material structures is small, conventional methods for determining thresholds used in binarization may not be sufficient to properly distinguish those structures. This is especially true when the brightness histogram shows a unimodal distribution, making threshold setting particularly difficult.

[0005] The problem that this invention aims to solve is to provide a material structure identification method that can accurately distinguish between multiple types of material structures in a microstructure observation image based on their brightness. [Means for solving the problem]

[0006] To solve the above problems, the material structure identification method according to the present invention has the following configuration. (1) In the material structure identification method according to the present invention, in a grayscale display of a structure observation image containing multiple types of material structures, the following steps are performed: a representative value extraction step to obtain the mode and half-width of the unprocessed histogram from the brightness histogram of the structure observation image; a normal distribution creation step to create a normal distribution having the mode and the half-width; a difference extraction step to extract a difference distribution by subtracting the normal distribution from the unprocessed histogram; a threshold determination step to set at least one brightness value corresponding to the intersection of the normal distribution and the difference distribution as a threshold; and a structure separation step to classify the material structures in the unprocessed histogram, with the threshold as the boundary, considering the high-brightness region and the low-brightness region as contributions to different material structures.

[0007] (2) In the embodiment of (1) above, the tissue observation image includes two types of material tissue, and each of the steps of the representative value extraction step, normal distribution creation step, difference extraction step, threshold determination step, and tissue separation step is performed once, and in the tissue separation step, the tissue observation image is binarized with the threshold as the boundary.

[0008] (3) In the embodiment of (1) or (2) above, the microstructure observation image is an optical microscope image obtained for iron or an iron-based alloy containing a martensite phase and a ferrite phase as the material structure, wherein the ferrite phase gives higher brightness than the martensite phase, and each of the representative value extraction step, normal distribution creation step, difference extraction step, threshold determination step, and microstructure separation step is performed once, and in the threshold determination step, the brightness value corresponding to the intersection that appears on the brightness side of the most frequent value is set as the threshold, and in the microstructure separation step, with the threshold as the boundary, the region on the low brightness side is considered to be the contribution of the martensite phase and the region on the high brightness side is considered to be the contribution of the ferrite phase.

[0009] (4) In the embodiment of (1) above, the microstructure observation image includes three or more types of material structures, and in the threshold determination step, if multiple intersections are obtained, the brightness value corresponding to each of the multiple intersections is used as the threshold, and in the microstructure separation step, the material structures are separated using each of the multiple thresholds as a boundary.

[0010] (5) In the embodiment of (1) or (4) above, the tissue observation image includes three or more types of material tissue, and while considering the continuous region of the difference distribution obtained in the difference extraction step as a new unprocessed histogram, the cycle of the representative value extraction step, the normal distribution creation step, and the difference extraction step is repeated multiple times to create multiple normal distributions, and in the threshold determination step, the brightness values ​​corresponding to the intersections between the multiple normal distributions and the intersections between the normal distribution and the difference distribution obtained in the last cycle are set as thresholds, thereby determining multiple thresholds, and in the tissue separation step, the material tissue is divided using each of the multiple thresholds as a boundary. [Effects of the Invention]

[0011] In the material microstructure identification method according to the invention described in (1) above, in the representative value extraction step, the mode and half-width of the unprocessed histogram are determined, and a normal distribution is created in the normal distribution creation step using these values. Since the brightness distribution of each material microstructure often takes the form of a normal distribution or a distribution close to it, the created normal distribution is subtracted from the unprocessed histogram in the difference extraction step to obtain a difference distribution, thereby distinguishing between material microstructures approximated by the created normal distribution and material microstructures approximated by the difference distribution. Then, by using the brightness corresponding to the intersection of the normal distribution and the difference distribution as a threshold, the material microstructures are divided into low brightness and high brightness sides, thereby enabling high-precision identification of material microstructures approximated by the normal distribution and material microstructures approximated by the difference distribution. Note that the difference distribution may include contributions from multiple material microstructures.

[0012] When using this method, the extraction of the mode and full width at half maximum in the representative value extraction process, the process of creating a normal distribution using the extraction results, the difference extraction process, and the threshold determination process can all be carried out without any arbitrariness. Therefore, even when the luminance histogram is unimodal or when the difference in luminance between multiple material structures is small, a clear and unambiguous threshold for distinguishing those material structures can be obtained. Furthermore, as in the case of analyses such as curve fitting, differences in analysts or the algorithms used do not affect the setting of the threshold. Whether or not a normal distribution or a distribution close to it appears in a part of the luminance histogram, such as on one side of the mode, does not affect whether or not a threshold can be set. Because there is no arbitrariness in each process, it is possible to automate the identification of material structures, including the determination of thresholds, in microstructure observation images, and even when identifying material structures for a large number of microstructure observation images, it is possible to eliminate the variation in identification and perform high-precision identification of structures.

[0013] In the embodiment described in (2) above, when the microstructure observation image contains two types of material structures, the microstructure observation image is binarized using a threshold value set by performing each step once. In this case, by setting a threshold value, the identification of the two structures by binarization can be performed with high accuracy. The obtained binarized image can be effectively used for analysis of the distribution of the two types of material structures, etc.

[0014] In the embodiment described in (3) above, when the microstructure observation image is an optical microscope image of iron or an iron-based alloy that includes a martensite phase and a ferrite phase, the two phases are distinguished using a threshold set to a higher brightness value than the mode. In optical microscope images, the martensite phase and the ferrite phase often give close brightness values, and in such cases, it is difficult to appropriately set a brightness threshold that can distinguish the two phases using conventional methods. However, according to the method of the present invention, as will be shown in later embodiments, the two phases can be distinguished with high accuracy using the threshold as a boundary.

[0015] In the embodiment of (4) above, when the microstructure observation image includes three or more material structures, and multiple predetermined intersections are obtained in the threshold determination step, the brightness values ​​corresponding to each of these intersections are used as thresholds to classify the material structures. When multiple intersections between the normal distribution and the difference distribution are obtained on both sides of the mode, etc., the brightness values ​​corresponding to these multiple intersections can be suitably adopted as thresholds for identifying three or more material structures. By setting thresholds in this way, even if the microstructure observation image is complex and includes many material structures, it becomes possible to identify each material structure and analyze its distribution.

[0016] In the embodiment of (5) above, when the microstructure observation image includes three or more material structures, the difference distribution is treated as a new unprocessed histogram, and multiple thresholds are determined by repeating the cycle for setting thresholds multiple times, and the material structures are divided using each of these multiple thresholds as a boundary. The difference distribution may be obtained as a composite of contributions from multiple material structures with small differences in brightness, but even in such cases, by repeating the above cycle multiple times, it becomes possible to separate the contributions of these multiple material structures and identify them. Therefore, even if the microstructure observation image is complex in which multiple material structures that give close brightness coexist, it becomes possible to identify each material structure and analyze its distribution. [Brief explanation of the drawing]

[0017] [Figure 1] This figure illustrates a method for identifying material microstructure according to one embodiment of the present invention. When the microstructure observation image includes two types of material microstructures, (a) shows the untreated histogram, (b) shows the normal distribution, and (c) shows the normal distribution and the difference distribution. [Figure 2] For cases where the tissue observation image includes three or more material tissues, (a) shows the untreated histogram, (b) shows the normal distribution, and (c) shows the normal distribution and the difference distribution. [Figure 3] In the case where the first difference distribution that appeared on the high-luminance side of Figure 2 is further separated by contribution, (a) shows an enlarged view of the first difference distribution, and (b) shows the state in which the first difference distribution is separated into the final normal distribution and the final difference distribution. (c) shows the distribution curves obtained throughout the entire process. [Figure 4] This diagram shows the steps involved in identifying the microstructure of an iron-based alloy using an optical microscope image. (a) is the microscope image, (b) is a binarized image obtained by visually identifying two types of microstructures, (c) is a binarized image obtained using a conventional method, and (d) is binarized using a new method. Additionally, (e) is a histogram of the microscope image, and (f) is a histogram separated by contribution. [Figure 5]This is a diagram showing each step when performing a similar analysis on a microscope image different from FIG. 4. The correspondence between the figure numbers and the image types is the same as that in FIG. 4. [Figure 6] This is a diagram showing each step when performing a similar analysis on yet another microscope image. (a) shows the microscope image, (b) shows the binarized image by the conventional method, and (c) shows the binarization by the new method. Also, (d) shows the histogram of the microscope image, and (e) shows the result of separating the contributions of the histogram. [Figure 7] This is a diagram showing each step when performing a similar analysis on yet another microscope image. The correspondence between the figure numbers and the image types is the same as that in FIG. 6.

Mode for Carrying Out the Invention

[0018] Hereinafter, a method for identifying a material structure according to an embodiment of the present invention will be described with reference to the drawings.

[0019] [Outline of the Material Structure Identification Method] In the material structure identification method according to this embodiment, a tissue observation image displayed in gradation, such as a grayscale image, including a plurality of types of material structures is taken as an analysis target, and a plurality of types of material structures are identified in the tissue observation image. That is, a plurality of types of material structures included in the tissue observation image are identified from each other, and the regions occupied by each target tissue are divided. For the identification of the tissue, a luminance threshold value is used, and the tissue is distinguished between the high-luminance side and the low-luminance side with the threshold value as a boundary.

[0020] The microstructure observation image targeted for microstructure identification is not particularly limited, as long as it includes regions corresponding to multiple types of material microstructures. The multiple types of material microstructures can be those that give different brightness levels on the microstructure observation image. Furthermore, any type of observation image is acceptable as long as the differences in material microstructures can be distinguished by brightness; examples include photographic images and various microscopic observation images from optical microscopes and electron microscopes. The number of coexisting material microstructures is also not particularly limited, as long as it is two or more types. The material species of each material microstructure is also not particularly limited, and various material microstructures of different material species, such as metals and inorganic compounds including metal oxides, can be targeted. Moreover, the coexisting multiple types of material microstructures may be of the same or different material species, but including at least some of the same type tends to reduce the brightness difference between the different material microstructures of the same material species, thus increasing the effectiveness of identifying material microstructures using the material microstructure identification method according to this embodiment. For example, a microstructure observation image containing multiple metallic microstructures can be a suitable target for identification.

[0021] In the material structure identification method according to this embodiment, a luminance histogram is created for the observed structure image, and a threshold is determined by performing calculations on the luminance histogram. Then, multiple structures included in the observed structure image are identified based on the determined threshold. Furthermore, based on the identification result, if necessary, image display is performed to distinguish the material structures, such as by binarization. The calculation process for determining the threshold is carried out in the following order: representative value extraction step, normal distribution creation step, difference extraction step, and threshold determination step. The cycle including the representative value extraction step, normal distribution creation step, and difference extraction step may be carried out multiple times. A tissue separation step is performed to identify the structures based on the determined threshold.

[0022] [Example involving two types of material structures] Here, the method for identifying the material structure according to this embodiment will be described in detail, using the case where the microstructure observation image includes two types of material structures as an example. As described above, the microstructure observation image that is the target of material structure identification is a grayscale or other graded display image. The microstructure observation image includes regions derived from two types of material structures that give different brightness, and preferably includes only two types of material structures, excluding unavoidable components. As a grayscale image including two types of material structures, for example, consider the image shown in Figure 4(a). The image shown in Figure 4(a) is an observation of an iron-based alloy containing a martensite phase and a ferrite phase after acid corrosion, and the martensite phase is observed to be dark and the ferrite phase is observed to be bright.

[0023] Prior to implementing the material microstructure identification method according to this embodiment, a luminance histogram is created from a microstructure observation image as illustrated in Figure 4(a). As illustrated in Figure 1(a), the luminance histogram is plotted by counting the pixels included in the microstructure observation image for each luminance value (grayscale value), with luminance displayed on the horizontal axis and frequency (number of pixels giving each luminance) displayed on the vertical axis. The luminance histogram shown in Figure 1(a) shows a nearly unimodal distribution with a gentle shoulder structure on the high-luminance side. In the material microstructure identification method according to this embodiment, the luminance histogram created from the microstructure observation image is referred to as the unprocessed histogram f(x).

[0024] (1) Representative value extraction process In the representative value extraction process, the mode μ and the full width at half maximum (FWHM) w are extracted as representative values ​​from the unprocessed histogram f(x). As shown in Figure 1(a), the mode μ refers to the brightness value at the point in the unprocessed histogram f(x) where the frequency value is the largest. In the case of a unimodal histogram, the brightness value corresponding to the peak of the unimodal histogram is the mode μ. The full width at half maximum (FWHM) w, also called the total width at half maximum, refers to the width of the unprocessed histogram f(x) in the region where the frequency value is half of the value (A) at the mode μ. In other words, the difference between the brightness value that gives a frequency of A / 2 on the brightness side of the unprocessed histogram f(x) above the mode μ and the brightness value that gives a frequency of A / 2 on the brightness side of the mode μ is the FWHM w. If the unprocessed histogram f(x) contains noise due to non-essential contributions, or if it is difficult to extract the mode μ and half-width w directly from the unprocessed histogram f(x), the unprocessed histogram f(x) may be subjected to appropriate processing such as smoothing before extracting the mode μ and half-width w. Furthermore, interpolation of data points may be performed as appropriate when extracting the mode μ and half-width w.

[0025] (2) Process for creating a normal distribution Next, in the normal distribution creation step, a normal distribution g(x) is created using the mode μ and half-width w obtained in the representative value extraction step. That is, as shown in Figure 1(b), a normal distribution g(x) is created that has the same mode μ and half-width w as those extracted from the untreated histogram f(x) in the representative value extraction step. Here, by the central limit theorem, the normal distribution g(x) can be considered to approximate the brightness distribution of one of the two material structures included in the microstructure observation image, typically the material structure with the larger occupied area. In the example in Figure 4(a), the normal distribution g(x) can be considered to approximate the distribution of the martensite phase.

[0026] (3) Difference extraction process In the next difference extraction step, as shown in Figure 1(c), the difference distribution d(x) is extracted by subtracting the normal distribution g(x) created in the normal distribution creation step from the unprocessed histogram f(x). That is, d(x) = f(x) - g(x). The obtained difference distribution d(x) can be considered to represent the brightness distribution of the material structures included in the microstructure observation image that are not approximated by the normal distribution g(x). If the microstructure observation image contains only two types of material structures, the difference distribution d(x) will represent the other type of material structure that is not approximated by the normal distribution g(x), and this difference distribution d(x) will take the form of a normal distribution or a distribution close to it, as shown in Figure 1(c).

[0027] (4) Threshold determination process Next, a threshold determination step is performed. Here, as shown in Figure 1(c), at least one luminance value corresponding to the intersection point c of the normal distribution g(x) created in the normal distribution creation step and the difference distribution d(x) obtained in the difference extraction step is set as the threshold T. As described above, the normal distribution g(x) and the difference distribution d(x) represent the luminance distributions of different material structures contained in the material structure, respectively, and the luminance at the intersection point c of the two distribution curves g(x) and d(x) becomes a representative value that distinguishes the luminance given by the material structures represented by each from one another.

[0028] Here, when the observed tissue image essentially contains only two types of material structures, the differential distribution d(x) often takes the form of a unimodal distribution close to a normal distribution. In such cases, where the differential distribution d(x) is obtained as a unimodal distribution, the luminance value corresponding to the intersection c of the simple differential distribution d(x) and the normal distribution g(x) can be used as the threshold T. On the other hand, when the differential distribution d(x) does not take such a simple shape and has multiple intersections with the normal distribution g(x), attention should be focused on the intersection located between the luminances given by the two material structures that should be distinguished from each other, and the luminance value corresponding to that intersection should be adopted as the threshold T. For example, if the microstructure observation image includes three or more material structures, the differential distribution d(x) (=d1(x)+d2(x)) may have two separate regions, a high-brightness side and a low-brightness side, as shown by the dashed line in Figure 2(c). As intersections with the normal distribution g(x), two intersections occur with respect to the vertex of the normal distribution g(x): intersection c1 on the high-brightness side and intersection c2 on the low-brightness side. In this case, it is necessary to distinguish the material structure that gives the highest brightness from the other material structures in the microstructure observation image, but it is not necessary to distinguish the material structure that gives the lowest brightness from the material structures that give higher brightness. In this case, the brightness T1 corresponding to the high-brightness intersection c1 can be adopted as the threshold T. Conversely, if it is necessary to distinguish the material structure that gives the lowest brightness from the other material structures, but it is not necessary to distinguish the material structure that gives the highest brightness from the material structures that give lower brightness. In this case, the brightness T2 corresponding to the low-brightness intersection c2 can be adopted as the threshold T.

[0029] (5) Tissue separation process Finally, in the tissue separation process, the material tissue is divided using the threshold T obtained in the threshold determination process described above. Specifically, using the threshold T as a boundary, the regions with higher brightness than the threshold T and the regions with lower brightness than the threshold T are considered to be contributions of different material tissues, thereby distinguishing multiple material tissues from one another. This makes it possible to clarify which of the multiple material tissues occupies each region in the microstructure observation image. Furthermore, by appropriately performing a binarization process with the threshold T as a boundary, the spatial distribution of each identified tissue is displayed as image information on the microstructure observation image, making it easier to visualize the distribution of each material tissue. In the case of the microstructure observation image exemplified in Figure 4(a), the regions showing brightness on the brightness side of the threshold can be identified as being occupied by the ferrite phase, and the regions showing brightness on the brightness side of the threshold can be identified as being occupied by the martensite phase. As will be explained in detail in later examples, Figure 4(d) shows the tissue observation image of Figure 4(a) binarized using the method described above. The martensite phase is shown in black and the ferrite phase in white, and both tissues are clearly distinguishable.

[0030] As described above, in the material microstructure identification method according to this embodiment, the mode μ and half-width w are extracted from the untreated histogram f(x), and a luminance threshold T for distinguishing multiple material microstructures from each other is set by a series of calculations using these values. If the series of calculations proceeds successfully, a threshold can be obtained. Therefore, even when it is difficult to determine an appropriate threshold by visual inspection of the histogram or by conventional analysis methods, such as when the difference in luminance given by multiple material microstructures in the microstructure observation image is small, or when the luminance histogram shows a unimodal distribution, the material microstructure identification method according to this embodiment can be applied to clearly determine the threshold. As shown in later examples, by using the threshold determined by the method according to this embodiment, material microstructures can be identified with higher accuracy compared to when using conventional methods.

[0031] If the luminance histogram can be approximated by a normal distribution on one side with the mode μ as the boundary, then a normal distribution function obtained by folding that side back to the other side can be used as an approximation of one type of material structure, instead of the normal distribution g(x) described above. Subtracting this normal distribution function from the luminance histogram is then considered as the contribution of another material structure. However, this method is only applicable when the distribution shape on one side of the mode μ in the luminance histogram can be considered as a normal distribution. In contrast, in the material structure identification method according to the embodiment described above, if the mode μ and half-width w can be extracted from the unprocessed histogram f(x), then regardless of whether or not a region that can be approximated by a normal distribution, such as on one side of the mode μ, is included in the unprocessed histogram f(x), the material structure can be identified by setting a threshold T using the same calculation process.

[0032] Furthermore, in the material structure identification method according to this embodiment, there is no arbitrariness in the extraction of the mode μ and half-width w in the representative value extraction step, and in the calculations in each subsequent step. Therefore, when there are multiple intersection points of the normal distribution g(x) and the difference distribution d(x), in the threshold determination step, if the only thing to decide is which intersection point to adopt for threshold determination, based on prior knowledge regarding the relationship between material structure type and brightness, then the threshold T for material structure identification can be uniquely determined. Unlike when regression analysis such as curve fitting is performed on a brightness histogram, the values ​​of the parameters used do not change depending on the discretion of the analyst or the characteristics of the algorithm used, so the threshold does not fluctuate due to non-essential influences such as differences in analysts or algorithms. Since each step can be performed without arbitrariness, it is possible to automate the identification of material structures in microstructure observation images, including the setting of the threshold. Even when identifying material structures for a large number of microstructure observation images, the material structure identification method according to this embodiment can be applied automatically to easily identify material structures and eliminate fluctuations in the identification results for each microstructure observation image.

[0033] As described above, the material structure identification method according to this embodiment is applicable regardless of the specific type of material structure, and the brightness threshold for identifying the material structure can be determined without uncertainty. However, corresponding to approximating the contribution of one type of material structure with a normal distribution g(x) to which the mode μ and half-width w of the unprocessed histogram f(x) are applied, the identification of the material structure using the set threshold can be performed with higher accuracy when one type of material structure occupies a large area in the microstructure observation image, and other material structures occupy a relatively small area. For example, it is preferable that the area occupied by one type of material structure in the microstructure observation image is 80% or more, and even more preferably 90% or more, of the total area of ​​the image.

[0034] [Example of a case involving three or more material structures] Up to this point, the material structure identification method according to this embodiment has been described using the case where the microstructure observation image includes two types of material structures as an example. However, the material structure identification method according to this embodiment can be similarly applied to cases where the microstructure observation image includes three or more types of material structures, and these three or more types of material structures can be identified from each other.

[0035] If the microstructure observation image contains essentially only two types of material structures, or if it contains three or more types of material structures but only needs to be classified into two types, then, as explained above, the material structures can be classified into two types by a single threshold T by performing each of the above steps once. On the other hand, if the microstructure observation image contains three or more types of material structures and these three or more types of material structures need to be classified from one another, then multiple thresholds can be set by applying at least one of the following methods (i) or (ii). Then, the material structures can be classified using each of these multiple thresholds as a boundary. (i) The representative value extraction process, the normal distribution creation process, and the difference extraction process are each performed once, and in the threshold determination process, multiple thresholds are set using one normal distribution. (ii) The cycle of representative value extraction, normal distribution creation, and difference extraction is repeated multiple times to obtain multiple normal distributions, and then a threshold determination process is carried out using these multiple normal distributions to set multiple thresholds. The following is a brief explanation of methods (i) and (ii).

[0036] First, let's briefly explain method (i) based on Figure 2. The unprocessed histogram f(x) shown in Figure 2(a) has a distribution that is tailed towards the low-luminance side compared to the histogram in Figure 1(a). Correspondingly, when we subtract the normal distribution g(x) in Figure 2(b), which was created from the mode μ and the half-width w, from the unprocessed histogram f(x), the resulting difference distribution d(x) has two separate regions, one on the high-luminance side and the other on the low-luminance side, as shown by the dashed line in Figure 2(c). In other words, the difference distribution d(x) has two regions that show a continuous, bell-shaped distribution, one on the high-luminance side and the other on the low-luminance side, with the mode μ in between. Here, we separate these two regions and designate the high-luminance region as the first difference distribution d1(x) and the low-luminance region as the second difference distribution d2(x).

[0037] As shown in Figure 2(c), the difference distribution d(x) has two intersection points c1 and c2 with the normal distribution g(x). That is, the first difference distribution d1(x) and the second difference distribution d2(x) each have intersection points c1 and c2 with the normal distribution g(x). Here, the luminance values ​​corresponding to these two intersection points c1 and c2 are defined as thresholds T1 and T2. In other words, the luminance value corresponding to the intersection point c1 of the first difference distribution d1(x) and the normal distribution g(x) is defined as the first threshold T1, and the luminance value corresponding to the intersection point c2 of the second difference distribution d2(x) and the normal distribution g(x) is defined as the second threshold T2. Here, T1 > T2. In this way, multiple thresholds can be determined in a single threshold determination process.

[0038] If multiple thresholds are obtained, in the tissue separation process, each threshold can be used as a boundary to separate the low-luminance region from the high-luminance region, assuming that each region is due to the contribution of a different material structure. As described above, if a first threshold T1 and a second threshold T2 are obtained, the region can be separated into three areas based on the contribution of the material structure: the high-luminance side of the first threshold T1, the luminance between the first threshold T2 and the second threshold T2, and the low-luminance side of the second threshold T2.

[0039] Figure 3(a) shows an enlarged view of the first differential distribution d1(x). This first differential distribution d1(x) does not have a shape close to a normal distribution, but rather has a shape that is tailed towards the high-luminance side. In other words, it is suggested that the first differential distribution d1(x) is not the result of only one type of material structure, but rather a composite of contributions from two or more types of material structures. In such cases, the method described in (ii) above can be applied to determine multiple thresholds from the first differential distribution d1(x).

[0040] In method (ii), the cycle of representative value extraction, normal distribution creation, and difference extraction is repeated multiple times, with continuous regions of the difference distribution obtained in the difference extraction step being treated as new unprocessed histograms. Then, multiple thresholds are determined by using the luminance values ​​corresponding to the intersections between multiple normal distributions obtained by repeating the cycle multiple times, and the intersections between the normal distribution and the difference distribution obtained in the last cycle, as thresholds.

[0041] In the example shown in Figure 3(a), the difference distribution d(x) obtained by performing the cycle of representative value extraction, normal distribution creation, and difference extraction once is considered as the first difference distribution d1(x), which is one of the continuous regions, and is treated as the new unprocessed histogram f(x). Here, a continuous region refers to a region in which the frequency values ​​are continuously positive. Then, the mode and half-width are extracted from this new unprocessed histogram f(x) using the representative value extraction process, and then, as shown in Figure 3(b), a normal distribution is created using the normal distribution creation process and a difference distribution is extracted using the difference extraction process. The normal distribution obtained here is taken as the final normal distribution g1(x), and the difference distribution is taken as the final difference distribution d1a(x).

[0042] In FIG. 3(c), the respective distribution curves obtained in the above analysis process, that is, the normal distribution g(x) created in the first normal distribution creation step, the second difference distribution d2(x) which is a part of the difference distribution obtained in the first difference extraction step, and the final normal distribution g1(x) and the final difference distribution d1a(x) obtained from the first difference distribution d1(x) in the second difference extraction step are shown together. In the threshold determination step, the luminance corresponding to the intersection point c1’ between the normal distribution g(x) obtained in the first cycle and the final normal distribution g1(x) obtained in the second cycle is defined as the threshold T1’. Further, the luminance corresponding to the intersection point c3 between the final normal distribution g1(x) and the final difference distribution d1a(x) obtained in the second cycle is defined as the threshold T3 (see FIG. 3(b)). In addition to the two thresholds T1’ and T3 thus defined, a threshold T2 is obtained, which is defined as the luminance value corresponding to the intersection point c2 between the second difference distribution d2(x) obtained as a part of the difference distribution d(x) in the first cycle and not being the target of the processing in the second cycle, and the normal distribution g(x) obtained in the first cycle. Thus, three thresholds T2, T1’, and T3 are obtained (T2 < T1’ < T3). Finally, in the tissue separation step, using each of those three thresholds T2, T1’, and T3 as a boundary, the material tissue is classified in the tissue observation image. That is, it is classified into four types of material tissues according to the luminance.

[0043] Here, the cycle of representative value extraction, normal distribution creation, and difference extraction was performed twice, but it may be performed three or more times. For example, in the configuration shown in Figure 3, the final difference distribution d1a(x), which is the difference distribution obtained in the second cycle, has a distribution close to a normal distribution. However, if the difference distribution d1a(x) obtained in this second cycle has a distribution shape far from a normal distribution, such as having a tail on the low-luminance side or the high-luminance side, this difference distribution may be treated as a new unprocessed histogram, and a third cycle may be performed. In this way, multiple cycles are repeated, and the normal distribution obtained in the last cycle is set as the final normal distribution, and the difference distribution is set as the final difference distribution. Then, in the threshold determination step, the luminance values ​​corresponding to the intersections between the multiple normal distributions obtained in each cycle, or more precisely, the intersections between the normal distributions obtained in the normal distribution creation steps of consecutive preceding and succeeding cycles, such as the first and second cycles, and the second and third cycles, are adopted as thresholds. In addition, the luminance values ​​corresponding to the intersections between the final normal distribution and the final difference distribution obtained in the last cycle are also adopted as thresholds. Furthermore, as shown in the second differential distribution d2(x) in Figure 3(c), if there is a differential distribution for which the above cycle has already been completed, the luminance value corresponding to the intersection of that differential distribution and the normal distribution used to extract that differential distribution (the normal distribution created in the cycle from which that differential distribution was obtained) is also adopted as a threshold.

[0044] Furthermore, if the difference distribution obtained in the difference extraction process in each cycle contains multiple consecutive regions that can be separated from each other, as shown in Figure 2(c) with the first difference distribution d1(x) and the second difference distribution d2(x), then the distribution curves of each of these regions may be treated as new unprocessed histograms, and the threshold setting for subsequent cycles may be carried out independently for each region. For example, in the configuration shown in Figure 2(c), the second difference distribution d2(x) takes a shape close to a normal distribution, but if this second difference distribution d2(x) takes a shape that deviates from a normal distribution, then the second or even subsequent cycles may be performed on the second difference distribution d2(x) independently of the first difference distribution d1(x). [Examples]

[0045] The present invention will be described in more detail below using examples. Here, we confirmed whether the accuracy of material structure identification can be improved by using the material structure identification method according to the embodiment of the present invention described above.

[0046] [Test Method] For iron-based alloys containing martensite and ferrite phases, samples were etched with Nital solution, and grayscale images were taken using an optical microscope to observe the microstructure. Four images were prepared for each of the four different samples.

[0047] A threshold for binarization was set in the brightness histogram of each tissue observation image. Two thresholds were set: a threshold Ta using a conventional discriminant identification method (Otsu's binarization method; hereafter referred to as the conventional method), and a threshold Tb using the method according to the embodiment of the present invention described above (hereafter referred to as the new method). Binarization of the tissue observation images was then performed by applying each threshold. In the new method, each of the following steps—representative value extraction, normal distribution creation, difference extraction, threshold determination, and tissue separation—was performed once. In the threshold determination step, the brightness value corresponding to the intersection of the normal distribution and the difference distribution, which appears on the brightness side of the mode, was set as the threshold Tb.

[0048] [Test Results] Figure 4 shows the steps involved in tissue identification for a single tissue observation image. Figure 4(a) is the tissue observation image obtained by microscopic observation. The relatively dark gray area that occupies most of the area in the image corresponds to the martensite phase, and the relatively bright scattered areas within the martensite phase, mainly distributed in the upper part of the image, correspond to the ferrite phase. Figure 4(b) is a binarized image in which the analyst visually identified the martensite phase and the ferrite phase, displaying the martensite phase in black and the ferrite phase in white. In this binarized image, the area percentage of the total area occupied by the ferrite phase (ferrite area percentage) is 3.4%.

[0049] Figure 4(e) shows the luminance histogram obtained from the tissue observation image in Figure 4(a). The histogram has a shape with a small shoulder-shaped subpeak on the high-luminance side of the main peak. Figure 4(f) shows the normal distribution g(x) obtained with the new method as a solid line, and the difference distribution d(x) as a dotted line. The difference distribution has two components, a low-luminance side and a high-luminance side, but we focus on the high-luminance side. The luminance value corresponding to the intersection of this component and the normal distribution is defined as the threshold Tb (indicated by an arrow). In the histogram in Figure 4(e), the threshold Tb is also displayed along with the threshold Ta from the conventional method. In a 256-level grayscale, the threshold Ta is 100 and the threshold Tb is 129, indicating a difference between the two.

[0050] Figures 4(c) and 4(d) show binarized images of the tissue observation image in Figure 4(a) using a threshold Ta from the conventional method and a threshold Tb from the new method, respectively. Each threshold acts as a boundary, with the low-luminance region belonging to the martensite phase and the high-luminance region belonging to the ferrite phase. Corresponding to the fact that threshold Ta is smaller than threshold Tb, Figure 4(c) has more white areas identifiable as the ferrite phase than Figure 4(d). Figure 4(c) does not accurately reflect the distribution of the ferrite phase in the tissue observation image in Figure 4(a) and the visually binarized image in Figure 4(b), and clearly shows an excessive amount of white areas identifiable as the ferrite phase. In contrast, Figure 4(d) shows a ferrite phase distribution very close to that of the tissue observation image in Figure 4(a) and the visually binarized image in Figure 4(b). The sparse distribution of the ferrite phase, mainly in the upper part of the image, is also well reproduced.

[0051] The comparison between binarized images obtained using the two thresholds described above and binarized images obtained by visual inspection can be quantitatively performed using the ferrite area ratio and the image matching rate calculated as the Dice coefficient. First, when using the conventional threshold shown in Figure 4(c), the ferrite area ratio is 60.1%, which is nearly 20 times the value of 3.4% in the visually inspected binarized image in Figure 4(b). The image matching rate is also low at 10.7%. In contrast, when using the threshold of the new method shown in Figure 4(d), the ferrite area ratio is 2.3%, which is close to the value of 3.4% in the visually inspected binarized image in Figure 4(b). The image matching rate is also significantly higher at 78.0% compared to the case in Figure 4(c). From these results, it can be confirmed that by setting a threshold using the new method and performing binarization of tissue observation images, a binarized image closer to that obtained when tissue is identified by human visual inspection is obtained than when using the conventional method, and that an appropriate threshold can be set.

[0052] Figure 5 shows the results of setting a threshold and performing binarization on a different tissue observation image in the same manner. In this case as well as in Figure 4 above, the binarized image using the new method with threshold Tb in Figure 5(d) reproduces the distribution of the ferrite phase in the tissue observation image in Figure 5(a) and the visually observed binarized image in Figure 5(b) much better than the binarized image using the conventional method with threshold Ta in Figure 5(c). The values ​​of the ferrite area ratio and image agreement rate described at the bottom of Figures 5(c) and (d) also confirm that the agreement rate with the visually observed binarized image is significantly higher in Figure 5(d).

[0053] Furthermore, Figures 6 and 7 show the results of setting thresholds and performing binarization on two other tissue observation images in the same manner. Visually binarized images are omitted here. Although the shape of the luminance histograms of these tissue observation images differs considerably from those in Figures 4 and 5, the binarized image using the new method of thresholding (c) remarkably reproduces the distribution of the ferrite phase in the tissue observation image (a) better than the binarized image using the conventional method of thresholding (b). These results demonstrate that by using the new method according to the embodiment of the present invention, tissue discrimination can be performed with high accuracy for tissue observation images that give luminance histograms of various shapes.

[0054] From the above analysis results, it was confirmed that by applying the material structure identification method according to an embodiment of the present invention to an optical microscope image obtained for iron or iron-based alloys containing a martensite phase and a ferrite phase as material structures, in which the ferrite phase exhibits higher brightness than the martensite phase, the image can be appropriately binarized. Specifically, by performing each of the representative value extraction step, normal distribution creation step, difference extraction step, threshold determination step, and structure separation step once, setting the brightness value corresponding to the intersection appearing on the brightness side of the mode in the threshold determination step as the threshold, and in the structure separation step, using that threshold as a boundary, considering the low-brightness region as the contribution of the martensite phase and the high-brightness region as the contribution of the ferrite phase, a binarized image that appropriately identifies the martensite phase and the ferrite phase can be obtained.

[0055] Although embodiments of the present invention have been described in detail above, the present invention is not limited to the above embodiments, and various modifications are possible without departing from the spirit of the invention.

Claims

1. In a grayscale microstructure observation image containing multiple types of material structures, For the brightness histogram of the aforementioned tissue observation image, A representative value extraction process to obtain the mode and full width at half maximum of the unprocessed histogram, A normal distribution creation step that creates a normal distribution having the mode and the half-width, A difference extraction step in which the difference distribution is extracted by subtracting the normal distribution from the unprocessed histogram, A threshold determination step is performed by a calculation processing means to determine a threshold value that is the brightness value corresponding to the intersection point of the normal distribution and the difference distribution, A method for identifying material microstructure, comprising: performing a microstructure separation step in which, in the untreated histogram, the high-luminance region and the low-luminance region are considered to be contributions of different material microstructures, with the threshold as the boundary, thereby classifying the material microstructure.

2. The aforementioned tissue observation image includes two types of material tissue, Each of the above steps—the representative value extraction step, the normal distribution creation step, the difference extraction step, the threshold determination step, and the tissue separation step—is performed once. The method for identifying a material structure according to claim 1, wherein, in the tissue separation step, the tissue observation image is binarized using the threshold as a boundary.

3. The aforementioned microstructural observation image is an optical microscope image obtained for iron or an iron-based alloy containing a martensite phase and a ferrite phase as material structure, and the ferrite phase exhibits higher brightness than the martensite phase. Each of the above steps—the representative value extraction step, the normal distribution creation step, the difference extraction step, the threshold determination step, and the tissue separation step—is performed once. In the threshold determination step, the luminance value corresponding to the intersection that appears on the luminance side of the mode is set as the threshold. The method for identifying the material structure according to claim 1, wherein, in the tissue separation step, the low-luminance region is considered to be the contribution of the martensite phase, and the high-luminance region is considered to be the contribution of the ferrite phase, with the threshold as the boundary.

4. The aforementioned tissue observation image includes three or more types of material tissue. In the threshold determination step, if multiple intersection points are obtained, the luminance value corresponding to each of the multiple intersection points is set as the threshold. The method for identifying a material structure according to claim 1, wherein in the tissue separation step, the material structure is divided using each of the plurality of thresholds as a boundary.

5. The aforementioned tissue observation image includes three or more types of material tissue. Of the difference distributions obtained in the difference extraction step, the continuous regions are newly considered as unprocessed histograms, By repeating the cycle of the aforementioned representative value extraction process, normal distribution creation process, and difference extraction process multiple times, Multiple normal distributions are created as described above. In the threshold determination step, multiple thresholds are determined by setting the luminance values ​​corresponding to the intersections between the multiple normal distributions and the intersections between the normal distribution and the difference distribution obtained in the last cycle as thresholds, respectively. The material structure identification method according to claim 1 or claim 4, wherein in the tissue separation step, the material structure is divided using each of the plurality of thresholds as a boundary.

Images