JP2019007944A - Segregation detection method and segregation detection device - Google Patents

Abstract

To detect a segregation area contained in a steel material cross-section with a high accuracy.SOLUTION: A segregation detection method acquires an image representing a transfer body of an etched steel material cross-section, and specifies a plurality of low brightness areas by a predetermined brightness in the image, calculates an image feature amount of each of the plurality of low brightness areas, creates classification data associated with a determination result of whether each of the plurality of low brightness areas is a segregation area by visual inspection and the image feature amount of each of the plurality of low brightness areas, defines the image feature amount contained in the classification data as an explanatory variable and defines the determination result as an objective variable, classifies whether each of a plurality of unknown low brightness areas is the segregation area by applying a model to the image feature amount of each of the plurality of unknown low brightness areas specified in the unknown image representing a transfer body for analysis after an analysis step of constructing the model of the explanatory variable for classifying the objective variable is completed, and detects each of the segregation areas classified as the segregation area.SELECTED DRAWING: Figure 3

Description

  The present invention relates to a technique for detecting a segregation region using an image representing a transfer body of an etched steel cross section.

  Conventionally, in order to evaluate the quality of steel materials, it is known to evaluate the area and size of segregation regions (hereinafter referred to as segregation regions) occurring in steel materials. In order to evaluate the area and size of this segregation region, it is known to perform component analysis of a steel material cross section by EPMA (Electron Probe Micro Analyzer). However, this component analysis has a problem in that the area of the segregation region to be evaluated is narrow and local, and it takes time to obtain the analysis result, so that feedback to the operation is delayed.

  Therefore, as a method for quickly evaluating the state of segregation, for example, as described in Patent Document 1, etc., image processing such as binarization is performed on an image representing a transfer body of an etched steel material cross section, A method of detecting a low luminance area as a segregation area from an image after image processing is known.

JP-A-1-167636

  However, in the above-described prior art, there is a possibility that the image after image processing includes an image representing a macro structure other than segregation, a processing mark at the time of etching, and the like as a low-luminance image similar to the segregation region. For this reason, an area other than the segregation area is excessively detected as a segregation area (hereinafter, overdetection), and the segregation area may not be detected with high accuracy.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a technique capable of detecting a segregation region included in a steel cross section with high accuracy.

  The segregation detection method according to one aspect of the present invention includes an image acquisition step of acquiring an image representing a transfer body of an etched steel cross section, and image processing for specifying a plurality of low-brightness regions having a lower brightness than a predetermined brightness in the image A step, an image feature amount calculating step for calculating an image feature amount for each of the plurality of low luminance regions, a determination result for visually determining whether each of the plurality of low luminance regions is a segregation region, and A classification data creating step for creating classification data in which the calculated image feature quantities of the plurality of low-luminance areas are associated with each other, and the image feature quantities included in the classification data are used as explanatory variables, and the classification A classification model construction step of constructing a classification model that is a model of the explanatory variable for classifying the objective variable, using the discrimination result included in the data as an objective variable, An analysis step to be executed, and after completion of the analysis step, the image acquisition step, the image processing step, and the image feature amount calculation using the transfer body for analysis different from the transfer body used in the analysis step. And sequentially executing the step, and applying the classification model to the image feature quantities of each of the plurality of unknown low-luminance regions specified in the unknown image representing the transfer body for analysis. A step of classifying whether each of the unknown low-intensity regions is a segregation region, and among the plurality of unknown low-intensity regions, each of the plurality of unknown low-intensity regions classified as a segregation region by the classification step And a segregation detection step for detecting as follows, and an analysis step for sequentially executing.

  In addition, the segregation detection apparatus according to an aspect of the present invention specifies an image acquisition unit that acquires an image representing a transfer body of an etched steel material cross section, and a plurality of low-brightness regions that have lower brightness than a predetermined brightness in the image. An image processing unit, an image feature amount calculating unit that calculates an image feature amount of each of the plurality of low-luminance regions, and a discrimination result that visually discriminates whether or not each of the plurality of low-luminance regions is a segregation region And a classification data creation unit that creates classification data that associates the calculated image feature amounts of the plurality of low-luminance regions with each other, and the image feature amount included in the classification data as an explanatory variable, A classification model construction unit that constructs a classification model that is a model of the explanatory variable for classifying the objective variable, the classification result included in the classification data as an objective variable, the image acquisition unit, and the image An analysis unit that performs an analysis process for sequentially operating a processing unit, the image feature amount calculation unit, the classification data creation unit, and the classification model construction unit, a classification unit, a segregation detection unit, and the analysis unit After completion of the analysis processing by the above, the image acquisition unit, the image processing unit, and the image feature amount calculation unit are operated in order using the transfer body for analysis different from the transfer body used in the analysis processing. And an analysis unit that performs an analysis process that sequentially operates the classification unit and the segregation detection unit. In the analysis process, the classification unit is identified in an unknown image that represents the transfer body for analysis. By applying the classification model to the image feature quantities of the plurality of unknown low-luminance regions, the plurality of unknown low-luminance regions are classified as segregation regions, and the segregation The detection unit is configured to detect the plurality of unread Of the low luminance area, a 其 people that have been classified as segregation area by the classifying unit, and detects as the segregation area.

  According to these configurations, the plurality of low-luminance areas are determined by using, as objective variables, the determination results obtained by visually determining whether or not each of the plurality of low-luminance areas specified in the image representing the transfer body is a segregation area. Using each image feature amount as an explanatory variable, a classification model that is a model of explanatory variables for classifying objective variables is constructed. Then, by applying the classification model to the image feature amount of each unknown low-brightness area specified in the unknown image representing the analysis transfer body, it is determined whether or not the unknown low-brightness area is a segregation area. being classified.

  For this reason, not only whether or not it is a low-luminance region as in the prior art, but also using the classification model, the image feature amount of the unknown low-luminance region is visually identified as a segregation region. It is possible to classify whether or not the unknown low luminance region is a segregation region with higher accuracy by further considering whether or not the image feature amount corresponds to the luminance region. As a result, the segregation region can be detected with high accuracy.

  Further, in the classification data creation step, the discrimination result and the image feature amount for each of the low luminance areas included in the predetermined first area of the image among the plurality of low luminance areas are associated with each other. The classification result and the image feature amount are associated with one classification data and the low luminance areas included in the second area different from the first area in the image among the plurality of low luminance areas. Second classification data individually, and in the classification model construction step, a first classification model is constructed using the first classification data, and a second classification model is constructed using the second classification data. In the classification step, among the plurality of unknown low-brightness areas, the image feature amount of each unknown low-brightness area included in the first corresponding area corresponding to the first area in the unknown image On the other hand, by applying the first classification model, to classify whether or not each unknown low luminance region included in the first corresponding region is a segregation region, among the plurality of unknown low luminance regions, Included in the second corresponding region by applying the second classification model to the image feature amount of each unknown low-luminance region included in the second corresponding region corresponding to the second region in the unknown image It is preferable to classify whether or not each unknown low-luminance region is a segregation region.

  According to this configuration, the first classification model based on the image feature amount of the low brightness area included in the first area in the image representing the transfer material, and the image feature amount of the low brightness area included in the second area of the image And a second classification model based on

  For this reason, in the classification step, the first classification model is applied to the image feature amount of each unknown low-luminance area included in the first corresponding area corresponding to the first area in the unknown image. It is possible to classify whether or not each of the unknown low-luminance areas is a segregation area with higher accuracy in consideration of an image feature amount specific to the low-luminance area that is discriminated as a segregation area by visual inspection. Similarly, each of the unknown low-luminance regions included in the second corresponding region is a segregation region in consideration of the image feature amount peculiar to the low-luminance region determined as a segregation region by visual observation included in the second region. It is possible to classify whether or not.

  The steel material cross section is a cross section when the slab is cut in parallel to the casting direction, and the first area is a central segregation analysis area extending in the horizontal direction in the vicinity of the vertical center in the image. Good.

  In the slab, in the vicinity of the center in the vertical direction, segregation, which is called center segregation, continues in a horizontal direction with respect to the casting direction. On the other hand, in addition to the center of the slab in the vertical direction, segregation called V segregation occurs obliquely with respect to the casting direction from the slab surface toward the center in the vertical direction.

  Therefore, according to this configuration, the center segregation in the unknown image is considered in consideration of the image feature amount peculiar to the low luminance area that is visually identified as the center segregation segregation area included in the center segregation analysis area that is the first area. It is possible to classify more accurately whether each unknown low-luminance region included in the first corresponding region corresponding to the analysis region is a segregation region of central segregation. Similarly, the unknown low luminance included in the second corresponding region is considered in consideration of the image features specific to the low luminance region included in the second region different from the central segregation analysis region and visually identified as the segregation region. Whether each region is a segregation region of V segregation can be classified with higher accuracy.

  Further, in the classification data creation step, among the plurality of low luminance regions, in the second region, the low luminance regions each included in a third region vertically above the central segregation analysis region. Included in the fourth region of the second region, which is the lower side of the center segregation analysis region in the second region, among the plurality of low luminance regions, the third classification data in which the discrimination result and the image feature amount are associated Fourth classification data in which the discrimination result and the image feature amount for each low-luminance region to be associated are individually created, and in the classification model construction step, the third classification data is used to generate a third classification data. Constructing a classification model, constructing a fourth classification model using the fourth classification data, and in the classification step, out of the plurality of unknown low-luminance regions, the second in the unknown image. By applying the third classification model to the image features of the unknown low-brightness areas included in the third corresponding areas corresponding to the areas, the unknown low-luminance areas included in the third corresponding areas, respectively. Is a segregation area, and among the plurality of unknown low-brightness areas, the image feature amount of each unknown low-brightness area included in the fourth corresponding area corresponding to the fourth area in the unknown image On the other hand, it is preferable to classify whether or not each unknown low-luminance region included in the fourth corresponding region is a segregation region by applying the fourth classification model.

  V segregation (hereinafter referred to as first V segregation) occurring in a region vertically above the center of the slab in the vertical direction and V segregation occurring in regions vertically below the center of the slab in the vertical direction ( Hereinafter, the second V segregation) has a characteristic in which the vertical center of the slab is inverted up and down in the vertical direction. That is, the image feature amount of the segregation region of the first V segregation obtained from the image representing the transfer material and the image feature amount of the segregation region of the second V segregation obtained from the image representing the transfer material of the cross section are: Have different characteristics.

  Therefore, according to the present configuration, the image feature amount peculiar to the low-luminance region, which is visually identified as the segregation region of the first V segregation, is included in the third region vertically above the center segregation analysis region. Thus, it is possible to classify more accurately whether or not each unknown low-luminance region included in the third corresponding region corresponding to the third region in the unknown image is a segregation region of the first V segregation. Similarly, in consideration of the image feature amount peculiar to the low luminance region, which is visually identified as the second V segregation segregation region, which is included in the fourth region vertically below the center segregation analysis region, Whether or not each unknown low-luminance region included in the four corresponding regions is a segregation region of the second V segregation can be classified with higher accuracy.

  Preferably, in the image processing step, binarization processing is performed on the image, and low-side luminance regions in the image after the binarization processing are specified as the plurality of low-luminance regions. .

  According to this configuration, the low-luminance area in the unknown image after binarization processing is identified as the unknown low-luminance area, and whether or not the identified unknown low-luminance area is a segregation area is classified with high accuracy. be able to.

  In the classification model construction step, it is preferable to construct a classification model using one or a plurality of statistical methods of decision trees, logistic regression, nearest neighbor methods, neural networks, support vector machines, and random forests.

  According to this configuration, using one or more statistical methods of decision tree, logistic regression, nearest neighbor method, neural network, support vector machine, and random forest used in general machine learning, A classification model can be built.

  Further, after the analysis step is completed, a derivation step for deriving a feature amount related to the segregation region of the unknown image based on the image feature amount of each segregation region detected in the segregation detection step, and visual inspection of the unknown image A determination data creating step for creating determination data in which the determination result for determining whether or not the quality of the unknown steel corresponding to the unknown image is good and the feature amount of the unknown image; and the determination data A determination model that builds a determination model that is a model of the explanatory variable for classifying the objective variable by using the feature amount included as an explanatory variable, the determination result included in the determination data as an objective variable, A quality analysis step for sequentially executing a construction step, and an evaluation pair different from the transcript for analysis after the quality analysis step is completed. Applying the determination model to the feature quantity of the unevaluated image representing the transfer body for the evaluation object after sequentially executing the analysis step and the derivation step using the transfer body for May further include a quality determination step of determining whether or not the quality of the unevaluated steel material corresponding to the unevaluated image is good.

  When determining whether the quality of the steel material is good or not only by visually observing the cross section of the steel material, the determination becomes personal and there is a possibility that the determination result may vary. In addition, it takes time to grasp the feature amount related to the segregation region of the steel material cross section, and it may not be possible to quickly determine whether or not the quality of the steel material is good.

  However, according to this configuration, the determination result that determines whether the quality of the unknown steel material is good or not by visual observation of the unknown image is the objective variable, and the characteristic amount related to the segregation region of the unknown image is the explanatory variable, and the objective variable is A determination model that is an explanatory variable model for classification is constructed. Thereafter, a derivation step is performed using the transfer body for evaluation, and a feature amount related to the segregated region of the unevaluated image representing the transfer body is derived. Then, by applying the determination model to the feature amount, it is determined whether or not the quality of the unevaluated steel material corresponding to the unevaluated image is good.

  For this reason, it is possible to quickly determine whether or not the quality of the unevaluated steel is good by using the feature amount derived by the derivation step without taking time and effort to grasp the feature amount regarding the segregation region of the unevaluated image. Can be determined. Further, not only the feature quantity related to the segregation area of the unevaluated image is considered, but also using the determination model, the feature quantity relates to the segregation area of the image corresponding to the steel material determined to have good quality in the past. It is possible to determine whether or not the quality of the unevaluated steel material is good with higher accuracy by further considering whether or not the feature amount is satisfied.

  In the deriving step, the feature amount of the predetermined first unknown region in the unknown image and the feature amount of the second unknown region different from the first unknown region in the unknown image are individually derived. In the determination data creation step, the determination result obtained by determining whether or not the quality of the first unknown steel material region corresponding to the first unknown region in the unknown steel material is good by visual observation of the first unknown region and the first Whether the quality of the second unknown steel region corresponding to the second unknown region in the unknown steel is good by visual inspection of the second unknown region and the first determination data in which the feature amount of the one unknown region is associated Second determination data in which the determination result for determining whether or not the feature amount of the second unknown region is associated with each other is created individually, and in the determination model construction step, the first determination A first judgment model is constructed using data, a second judgment model is constructed using the second judgment data, and in the quality judgment step, the first judgment model corresponding to the first unknown area in the unevaluated image is constructed. By applying the first determination model to the feature amount of the evaluation region, it is determined whether or not the quality of the first steel material region corresponding to the first unevaluated region in the unevaluated steel material is good. And applying the second determination model to the feature quantity of the second unevaluated area corresponding to the second unknown area in the unevaluated image, thereby providing the second unevaluated area in the unevaluated steel material. It may be determined whether or not the quality of the second steel material region corresponding to is good.

  According to this configuration, the first determination model based on the feature amount related to the segregation region of the first unknown region in the unknown image and the second determination model based on the feature amount related to the segregation region of the second unknown region in the unknown image include: Built individually.

  For this reason, in the quality determination step, by applying the first determination model to the feature amount related to the segregation region of the first unevaluated region corresponding to the first unknown region in the unevaluated image, In consideration of the characteristic amount related to the segregation region, it is possible to more accurately and quickly determine whether or not the quality of the first steel material region corresponding to the first unevaluated region in the unevaluated steel material is good. In the same manner, considering the feature amount related to the segregation region of the second unevaluated region, it is more accurate whether the quality of the second steel region corresponding to the second unevaluated region in the unevaluated steel is good or not. A good and quick decision can be made.

  The steel material cross section is a cross section when the slab is cut in parallel with the casting direction, and the first unknown area is a central segregation analysis area extending in the horizontal direction near the center in the vertical direction in the unknown image. May be.

  In the slab, in the vicinity of the center in the vertical direction, segregation, which is called center segregation, continues in a horizontal direction with respect to the casting direction. On the other hand, in addition to the center of the slab in the vertical direction, segregation called V segregation occurs obliquely with respect to the casting direction from the slab surface toward the center in the vertical direction.

  Therefore, according to the present configuration, in consideration of the feature amount related to the segregation region of the center segregation included in the first unevaluated region extending in the horizontal direction near the center in the vertical direction in the unevaluated image, Whether or not the quality of the first steel material region extending in the horizontal direction in the vicinity of the center is good can be determined more accurately and quickly. Similarly, in consideration of the feature amount related to the segregation region of V segregation included in the second unevaluated region different from the first unevaluated region in the unevaluated image, the first different region from the first steel region in the unevaluated steel material. Whether or not the quality of the two-steel material region is good can be determined more accurately and quickly.

  In the determination model construction step, it is preferable to construct a determination model using one or more statistical methods of decision tree, logistic regression, nearest neighbor method, neural network, support vector machine, and random forest.

  According to this configuration, using one or more statistical methods of decision tree, logistic regression, nearest neighbor method, neural network, support vector machine, and random forest used in general machine learning, A decision model can be constructed.

  ADVANTAGE OF THE INVENTION According to this invention, the segregation area | region contained in the steel material cross section can be detected with high precision.

It is a block diagram which shows the structure of a segregation detection apparatus. It is a figure which shows an example of a transfer image. It is a flowchart which shows an example of an analysis process. It is a flowchart which shows an example of an analysis process. It is a flowchart which shows another example of an analysis process. It is a flowchart which shows another example of an analysis process. It is a flowchart which shows another example of an analysis process. It is a flowchart which shows another example of an analysis process. It is a block diagram which shows another example of a structure of a control part. It is a flowchart which shows an example of a quality analysis process. It is a flowchart which shows an example of steel material quality determination processing. It is a flowchart which shows another example of a quality analysis process. It is a flowchart which shows another example of steel material quality determination processing.

(First embodiment)
Hereinafter, a first embodiment according to the present invention will be described. In addition, the structure which attached | subjected the same code | symbol in each figure shows that it is the same structure, The description is abbreviate | omitted suitably.

  FIG. 1 is a block diagram showing the configuration of the segregation detection apparatus 1. As shown in FIG. 1, the segregation detection apparatus 1 according to the embodiment of the present invention includes an imaging unit 30, an input unit 50, a display unit 60, a communication unit 70, a storage unit 90, a control unit 10, It has.

  The photographing unit 30 is configured by a camera, for example, and outputs an image obtained by photographing a subject to the control unit 10.

  The input unit 50 includes an input device such as a keyboard, a mouse, and a touch panel, and outputs various operation instructions and information input by the user using the input device to the control unit 10.

  The display unit 60 is configured by a display device such as a liquid crystal display, and displays instructed information under the control of the control unit 10.

  The communication unit 70 is connected to a network (not shown) and performs communication with an external device such as a personal computer connected to the network.

  The storage unit 90 is configured by a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive). Information used for control of the control unit 10 is stored in the storage unit 90 in advance. The storage unit 90 stores various types of information under the control of the control unit 10.

  The control unit 10 controls the entire segregation detection apparatus 1. Specifically, the control unit 10 includes a microcomputer provided with a CPU, RAM, ROM, and the like. When the CPU executes a control program stored in the ROM, the control unit 10 may, for example, include an image acquisition unit 101, an image processing unit 102, an image feature amount calculation unit 103, a classification data creation unit 104, and a classification model construction. Functions as a unit 105, an analysis unit 106, an analysis unit 107, a classification unit 108, and a segregation detection unit 109.

  The image acquisition unit 101 acquires an image representing a transfer body of an etched steel section. Specifically, the image acquisition unit 101 acquires an image output to the control unit 10 when the imaging unit 30 images the transfer body as a subject as an image representing the transfer body. Not limited to this, the image acquisition unit 101 may acquire an image representing the transfer body received by the communication unit 70 from an external device.

  Hereinafter, it is assumed that the steel material cross section is a cross section when a plate-shaped slab generated by continuous casting is cut in parallel to the casting direction. In addition, a transfer body having an etched steel cross section is produced, for example, as follows. After etching the steel section with a picric acid-based corrosive solution, it is washed with water and dried. Thereafter, a highly viscous liquid substance such as grease is applied to the steel cross section, the steel cross section is lightly polished with emery paper, and the highly viscous liquid substance on the surface of the steel cross section is wiped off. Subsequently, the highly viscous liquid material remaining in the corroded portion is transferred to a transparent thin film in which an adhesive is glued on one side such as cello tape (registered trademark). That is, the thin film is produced as the transfer body.

  The image processing unit 102 specifies a plurality of low-brightness areas having lower luminance than a predetermined luminance in an image representing the transfer body acquired by the image acquisition unit 101 (hereinafter, transfer image). Specifically, the image processing unit 102 performs binarization processing on the transferred image. The threshold value used in the binarization processing is such that the luminance values of all the areas determined to be segregation areas (hereinafter referred to as segregation areas) by visually observing a plurality of transfer images are lower on the lower side of the binary values. The brightness value is determined to be higher than the highest value of the brightness values of all the regions so as to be converted into the brightness value.

  Therefore, the image processing unit 102 performs a binarization process using the threshold value on the transfer image, thereby obtaining a luminance value lower than the threshold value included in the transfer image on the lower side of the binary values. Convert to luminance value. Then, the image processing unit 102, in the transfer image after the binarization processing, displays a plurality of regions whose luminance values are the lower luminance values as a plurality of luminances lower than the predetermined luminance represented by the threshold value. Identified as a low brightness area. The plurality of identified low-luminance regions may include not only segregation regions but also regions such as macrostructures other than segregation and processing marks during etching.

  Immediately before the image processing unit 102 performs binarization processing on the transfer image, the image quality of the transfer image is improved, for example, noise removal processing or density unevenness removal processing. For this purpose, known image processing may be performed.

  Specifically, the image processing unit 102 performs a known filter process using a smoothing filter, a median filter, an expansion filter, a contraction filter, a Gaussian filter, a double literal filter, or the like as a process for removing noise. Also good. Further, the image processing unit 102 may perform filter processing using a high-pass filter that excludes frequency components of a predetermined frequency or higher as processing for removing density unevenness. The predetermined frequency may be set to a frequency that is considered to most reduce density unevenness based on experimental values and the like. Further, immediately before the image processing unit 102 performs binarization processing on the transfer image, the transfer processing is performed on the transfer image using a bandpass filter that simultaneously removes noise and density unevenness. Also good. The threshold used when applying the bandpass filter may be set to a value that is considered to be able to most reduce noise and density unevenness based on experimental values and the like.

  Further, the method by which the image processing unit 102 specifies a plurality of low luminance areas in the transfer image is not limited to the above. For example, the image processing unit 102 may perform known image processing that directly specifies a plurality of regions having a luminance lower than a predetermined luminance in the transferred image without performing image processing such as binarization processing. .

  The image feature amount calculation unit 103 calculates an image feature amount for each of the plurality of low luminance regions specified by the image processing unit 102. The image feature amount includes, for example, an area, an average value of luminance values, a standard deviation of luminance values, a perimeter length, a rectangularity, and a roundness (a true value obtained from the maximum distance from the center) for each of a plurality of low luminance areas. Ratio of actual area to circle area), compactness (ratio of actual circumference to the circumference of a perfect circle obtained from the area), convexity, bulkiness ratio (ratio of actual area to approximate elliptical area), hole (only in the region) Number of hole-like (high luminance) regions), the total area of holes, and the like.

  FIG. 2 is a diagram illustrating an example of a transfer image. As shown in FIG. 2, the image feature amount includes, for example, information (for example, x1, y1) indicating the center position of each of a plurality of low-luminance areas and the vertical direction (height direction) in the transferred image. The distance (example: a1) and direction (example: downward) from the center position Yc to the center position of each of the plurality of low-luminance regions are included. The image feature amount includes a width (for example, w1) and a height (for example: h1) of a circumscribed rectangle parallel to the axis that circumscribes each of the plurality of low-luminance regions, and a width of the circumscribed rectangle (for example: w1) and height (example: h1) ratio (aspect ratio) (for example, w1: h1), width (example: wa1) and height (example: ha1) of a circumscribed rectangle parallel to any two axes May be. In addition, the image feature amount may include a width / height of the approximate ellipse of each of the plurality of low luminance regions, and an inclination angle θ with respect to the casting direction of the long axis.

  Returning to FIG. The classification data creation unit 104 visually determines whether or not each of the plurality of low-brightness areas specified by the image processing unit 102 is a segregation area, and the plurality of pieces calculated by the image feature amount calculation unit 103. Classification data in which the image feature amounts of the low-luminance regions are associated with each other is created.

  For example, the classification data creation unit 104 causes the display unit 60 to display a transfer image after the binarization processing by the image processing unit 102. The user who has visually observed the transfer image displayed on the display unit 60 uses the input unit 50 to select each of a plurality of low-luminance regions specified by the image processing unit 102 included in the transfer image. A determination result as to whether or not the selected low luminance region is a segregation region is input. The classification data creation unit 104 determines whether each of the plurality of input low-brightness areas is a segregation area and the plurality of low-brightness areas calculated by the image feature amount calculation unit 103. The classification data in which the image feature amounts are associated with each other is created.

  The classification model construction unit 105 uses the image feature amount included in the classification data created by the classification data creation unit 104 as an explanatory variable, uses the determination result contained in the classification data as an objective variable, and sets the objective variable as the objective variable. A model of the explanatory variables for classification (hereinafter referred to as a classification model) is constructed using a known statistical method.

  The statistical method includes, for example, an explanatory variable (a plurality of low-luminance regions) so that the purity of the objective variable (a discrimination result obtained by visually determining whether each of the plurality of low-luminance regions is a segregation region) is increased. Logistics that determine decision tree analysis that builds a rule (classification model) that combines tree features of each image feature) and its threshold value, and a regression equation (classification model) to a logistic curve that includes objective variables and explanatory variables Techniques used in general machine learning such as regression analysis can be applied. For the statistical method, for example, a method used in general machine learning such as nearest neighbor method, neural network, support vector machine, and random forest can be applied. Alternatively, the classification model construction unit 105 may construct a classification model using a plurality of methods among the statistical methods described above.

  The analysis unit 106 performs an analysis process (analysis step) that sequentially operates the image acquisition unit 101, the image processing unit 102, the image feature amount calculation unit 103, the classification data creation unit 104, and the classification model construction unit 105. Do.

  FIG. 3 is a flowchart illustrating an example of analysis processing. Specifically, when an analysis process start instruction is input by the input unit 50, the analysis unit 106 starts the analysis process. As shown in FIG. 3, when the analysis unit 106 starts the analysis process, first, the image acquisition unit 101 is operated to acquire an image (hereinafter referred to as a teaching image) representing a transfer body of the etched steel cross section (S11). (Image acquisition step)).

  Next, the analysis unit 106 operates the image processing unit 102 to specify a plurality of low-brightness areas having lower luminance than the predetermined luminance in the teaching image (S12 (image processing step)). Next, the analysis unit 106 operates the image feature amount calculation unit 103 to calculate the image feature amount of each of the plurality of low luminance regions specified in the teaching image (S13 (image feature amount calculation step)).

  Next, the analysis unit 106 operates the classification data creation unit 104 to determine whether or not each of the plurality of low-brightness areas specified in the teaching image is a segregation area, and the calculation result in S13. Classification data in which the image feature quantities of the plurality of low-luminance areas are associated with each other is created (S14 (classification data creation step)).

  Then, the analysis unit 106 operates the classification model construction unit 105 to use the image feature amount included in the classification data created in S14 as an explanatory variable, and the determination result included in the classification data as an objective variable. Then, a classification model that is a model of the explanatory variable for classifying the objective variable is constructed (S15 (classification model construction step)).

  Returning to FIG. After the analysis processing by the analysis unit 106 is completed, the analysis unit 107 operates the image acquisition unit 101, the image processing unit 102, and the image feature amount calculation unit 103 in order using an analysis transfer body, and then described later. An analysis process (analysis step) for sequentially operating the classification unit 108 and the segregation detection unit 109 is performed.

  Here, the transfer body for analysis is a transfer body that is different from the transfer body that is used when the analysis unit 106 operates the image acquisition unit 101 in the analysis process and that transfers the cross section of the steel material to be analyzed. Specifically, the transfer body for analysis is created by transferring the cross section of the steel material to be analyzed to the thin film after etching the cross section of the steel material to be analyzed, similarly to the transfer body having transferred the cross section of the steel material to be analyzed.

  FIG. 4 is a flowchart illustrating an example of the analysis process. Specifically, after the analysis process by the analysis unit 106 is completed, when an analysis process start instruction is input by the input unit 50, the analysis unit 107 starts the analysis process. As shown in FIG. 4, when the analysis unit 107 starts the analysis process, first, the image acquisition unit 101 is operated to acquire an image representing an analysis transfer body (hereinafter, unknown image) (S21 (image acquisition). Step)).

  Next, the analysis unit 107 operates the image processing unit 102 to specify a plurality of low luminance regions (hereinafter, unknown low luminance regions) having a luminance lower than a predetermined luminance in the unknown image (S22 (image processing step)). Next, the analysis unit 107 operates the image feature amount calculation unit 103 to calculate the image feature amount of each of the plurality of unknown low luminance areas specified in the unknown image (S23 (image feature amount calculation step)).

  The analysis unit 107 operates the classification unit 108 (S24 (classification step)), operates the segregation detection unit 109 (S25 (segregation detection step)), and ends the analysis process.

  Details of S24 and S25 will be described below. In the description, details of the classification unit 108 (FIG. 1) and the segregation detection unit 109 (FIG. 1) will be described.

  In S24, the classification unit 108 applies the classification model constructed in S15 (FIG. 3) to the image feature amounts of the plurality of unknown low-luminance regions calculated in S23, so that a plurality of unknown low-levels are obtained. It classifies whether each brightness | luminance area | region is a segregation area | region.

  For example, it is assumed that a classification model is constructed by a decision tree analysis method in S15 (FIG. 3). In other words, the explanatory variable (the image of each of the plurality of low-luminance regions is increased) so that the purity of the objective variable (the determination result obtained by visually determining whether or not each of the plurality of low-luminance regions is a segregation region) increases. It is assumed that the rule is a combination of a feature amount and a threshold value in a tree shape. In this case, in S24 (FIG. 4), the classification unit 108 determines the image feature amount included as the explanatory variable in the classification model among the image feature amounts of the plurality of unknown low-luminance regions, and the classification model. Whether or not each of the plurality of unknown low-luminance regions is a segregation region is classified according to the comparison result with the threshold value of the image feature amount.

  Alternatively, it is assumed that a classification model is constructed by a logistic regression analysis method in S15 (FIG. 3). That is, the classification model is an objective variable constructed by logistic regression analysis (discrimination result obtained by visually determining whether or not each of the plurality of low-luminance regions is a segregation region) and an explanatory variable (each of the plurality of low-luminance regions). It is assumed that this is a regression equation to a logistic curve including the image feature amount. In this case, in S24 (FIG. 4), the classification unit 108 selects the image feature amount included as the explanatory variable in the regression equation indicated by the classification model among the image feature amounts of the plurality of unknown low-luminance regions. Whether or not each of the plurality of unknown low-brightness areas is a segregation area is classified according to the result substituted into the regression equation.

  In S25 (FIG. 4), the segregation detection unit 109 detects each of the plurality of unknown low-luminance regions classified as the segregation region in S24 as a segregation region.

  Therefore, according to the configuration of the first embodiment, not only whether or not it is a low-luminance region as in the prior art, but also using the classification model, the image feature amount of the unknown low-luminance region is visually It is possible to classify whether or not the unknown low-brightness area is the segregation area with higher accuracy by further considering whether or not the image feature amount of the low-brightness area determined as the segregation area is obtained. As a result, the segregation region can be detected with high accuracy.

(Second embodiment)
Hereinafter, a second embodiment according to the present invention will be described. In the configuration of the first embodiment, classification data is created based on the image feature amounts of a plurality of low-luminance areas included in the teaching image, and a classification model is constructed using the classification data. And the segregation area | region was detected by classifying whether the some unknown low-intensity area | region contained in an unknown image is a segregation area | region using the said classification model.

  However, in the configuration of the second embodiment, unlike the configuration of the first embodiment, the second embodiment is based on the image feature amounts of the plurality of low luminance regions included in the two regions included in the teaching image. Classification data corresponding to each of the regions is created. Then, using the created two classification data, two classification models corresponding to the two areas are individually constructed. Whether or not each of the plurality of unknown low-luminance regions included in the two regions corresponding to the two regions in the unknown image is a segregation region using the two classification models constructed. By segregating, segregation regions are detected.

  Hereinafter, the reason for recalling the configuration of the second embodiment will be described. Segregation that is harmful to the quality of a plate-shaped slab produced by continuous casting is generally classified into center segregation and V segregation. As shown in FIG. 2, the center segregation is segregation that occurs in a row in the vicinity of the center position Yc in the vertical direction (height direction) of the cross section of the cast slab. V segregation is segregation that occurs continuously at an angle from the outer periphery of the slab toward the center position Yc.

  The tendency of the inclination angle θ is different between the center segregation segregation region and the V segregation segregation region seen in the transferred image. The inclination angle θ is an angle of the major axis of an ellipse approximating each segregation region with respect to the casting direction. For example, as schematically shown in FIG. 2, the inclination of the segregation region of the center segregation is almost horizontal, and the inclination angle θ is the largest near 0 degree. On the other hand, the segregation region of V segregation has an inclination, and the inclination angle θ tends to have a magnitude greater than a certain value.

  Therefore, as in the first embodiment, constructing a classification model of image feature amounts for classifying whether or not a segregation region exists without distinguishing between center segregation and V segregation complicates the construction of the classification model. This causes a decrease in segregation region detection accuracy. Therefore, an image feature amount classification model for classifying whether or not it is a segregation region of central segregation and an image feature amount classification model for classifying whether or not it is a segregation region of V segregation are separately constructed, Recalling that each of the two classification models that were constructed is used to individually classify whether the unknown low-brightness region is a segregation region representing central segregation or whether it is a segregation region representing V segregation. did.

  Specifically, in the configuration of the second embodiment, the analysis unit 106 performs analysis processing according to the flow shown in FIG. FIG. 5 is a flowchart illustrating another example of the analysis process.

  In the following description, as shown in FIG. 2, the horizontal direction (casting direction) within a predetermined distance d from the vertical center position Yc of the transferred image, which is considered to cause center segregation. A region extending in the direction (predetermined first region) is referred to as a center segregation analysis region.

  Further, as shown in FIG. 2, in a region (second region) different from the center segregation analysis region in the transfer image, where V segregation is considered to occur, a region (third region) above the center segregation analysis region in the vertical direction. (Region) is described as a first V segregation analysis region, and a region (fourth region) on the lower side in the vertical direction than the center segregation analysis region is described as a second V segregation analysis region.

  As shown in FIG. 5, when the analysis process is started by the analysis unit 106, S11 to S13 are performed as in the first embodiment. Next, the analysis unit 106 operates the classification data creation unit 104 to execute S31. In S31, the classification data creation unit 104 includes each of the plurality of low-luminance regions identified in S12 in the central segregation analysis region, the first V segregation analysis region, or the second V segregation analysis region. Is determined (S31).

  Specifically, in S31, as shown in FIG. 2, the classification data creation unit 104 starts from the center position Yc in the vertical direction (height direction) in the transferred image, which is included in the image feature amount calculated in S13. If the distance (for example, a1) to the center position of each of the plurality of low-luminance areas specified in S12 is within a predetermined distance d, each of the low-luminance areas becomes the central segregation analysis area. Judged to be included.

  On the other hand, the classification data creation unit 104 determines that the distance from the center position Yc to the center position of each of the plurality of low luminance areas specified in S12 is longer than the predetermined distance d and the image calculated in S13. When the direction from the center position Yc to the center position of each low-luminance area included in the feature value indicates upward, it is determined that each of the low-luminance areas is included in the first V segregation analysis area. .

  Further, the classification data creation unit 104 has a distance from the center position Yc to the center position of each of the plurality of low luminance areas specified in S12 is longer than the fixed distance d, and the center position Yc When the direction to the center position of each low-luminance area indicates downward, it is determined that each of the low-luminance areas is included in the second V segregation analysis area.

  Then, the classification data creation unit 104 performs S14a for the low-luminance region determined to be included in the central segregation analysis region in S31. Specifically, in S14a, similarly to S14 (FIG. 3), the classification data creation unit 104 determines whether or not each low-luminance region determined to be included in the central segregation analysis region in S31 is a segregation region. Central segregation classification data (first classification data) is created by associating the discrimination result determined by visual observation with the image feature amount of each low-luminance area calculated in S13 (S14a (classification data creation step)) ).

  In this case, the analysis unit 106 operates the classification model construction unit 105, and similarly to S15 (FIG. 3), the image feature amount included in the center segregation classification data created in S14a is used as an explanatory variable, and the center segregation is performed. A central segregation classification model (first classification model) is constructed using the discrimination results included in the classification data as objective variables (S15a (classification model construction step)).

  On the other hand, the classification data creation unit 104 performs S14b on the low-luminance region determined to be included in the first V segregation analysis region and the second V segregation analysis region (second region) in S31. However, the classification data creation unit 104 performs S32 before performing S14b when performing S14b for the low-luminance region determined to be included in the first V segregation analysis region in S31.

  Specifically, in S32, the classification data creation unit 104 calculates the image feature amount of each low-luminance area calculated in S13 and determined to be included in the first V segregation analysis area in S31 in the teaching image. The center position Yc (FIG. 2) in the direction is turned upside down (S32).

  For example, in S <b> 32, the classification data creation unit 104 determines the inclination of the major axis of the ellipse that approximates the low luminance region with respect to the casting direction, which is the image feature amount of the low luminance region determined to be included in the first V segregation analysis region. The angle θ (FIG. 2) is multiplied by −1. As a result, the image feature amount is inverted up and down with the vertical center position Yc (FIG. 2) in the teaching image as an axis.

  In S14b, similarly to S14 (FIG. 3), the classification data creation unit 104 visually checks whether or not each of the low luminance areas determined to be included in the first V segregation analysis area in S31 is a segregation area. Classification data in which the determined determination result is associated with the image feature amount of each low-luminance region after S32 is created. Further, similarly to S14 (FIG. 3), the classification data creation unit 104 visually determines whether or not each of the low-luminance regions determined to be included in the second V segregation analysis region in S31 is a segregation region. Classification data in which the determination result is associated with the image feature amount of each low-luminance area calculated in S13 is created. Then, the classification data creation unit 104 creates the two created classification data as V segregation classification data (second classification data) (S14b (classification data creation step)).

  In this case, the analysis unit 106 operates the classification model construction unit 105, and similarly to S15 (FIG. 3), the image feature amount included in the V segregation classification data created in S14b is used as an explanatory variable, and the V segregation is performed. A V segregation classification model (second classification model) is constructed using the discrimination result included in the classification data as an objective variable (S15b (classification model construction step)).

  Then, when the construction of the center segregation classification model in S15a and the construction of the V segregation classification model in S15b are completed, the analysis unit 106 ends the analysis process.

  On the other hand, in the configuration of the second embodiment, the analysis unit 107 executes analysis processing according to the flow shown in FIG. FIG. 6 is a flowchart illustrating another example of the analysis process.

  As shown in FIG. 6, when the analysis process is started by the analysis unit 107, S21 to S23 are performed as in the first embodiment. Next, the analysis unit 107 operates the classification unit 108 to execute S41. In S41, the classification unit 108 determines that each of the plurality of unknown low-luminance regions identified in S22 is a region corresponding to the center segregation analysis region (first correspondence region) or first V segregation analysis region in the unknown image. It is determined which of the corresponding region (third corresponding region) and the region corresponding to the second V segregation analysis region (fourth corresponding region) is included (S41).

  The region corresponding to the center segregation analysis region in the unknown image is a horizontal direction (casting direction) within a predetermined distance d from the vertical center position of the unknown image where the center segregation is considered to occur. ). The region corresponding to the first V segregation analysis region in the unknown image is different from the region corresponding to the center segregation analysis region in the unknown image where V segregation is considered to occur. Is also a region on the upper side in the vertical direction. The region corresponding to the second V segregation analysis region in the unknown image is different from the region corresponding to the center segregation analysis region in the unknown image where V segregation is considered to occur. Is also a region on the lower side in the vertical direction.

  Specifically, in S41, the classification unit 108 determines from the center position Yc in the vertical direction (height direction) in the unknown image included in the image feature amount calculated in S23, as in S31 (FIG. 5). Based on the distance and direction to the center position of each of the plurality of unknown low-luminance areas identified in (2), the unknown low-luminance areas each correspond to the central segregation analysis area, the areas corresponding to the first V segregation analysis area, and It is determined which of the regions corresponding to the second V segregation analysis region is included.

  Then, the classification unit 108 performs S24a on the unknown low-luminance area determined to be included in the area corresponding to the central segregation analysis area in S41. Specifically, in S24a, the classification unit 108 performs S15a (FIG. 5) on the image feature amount of each unknown low-luminance area of S24a calculated in S23, as in S24 (FIG. 4). ) Is applied to classify whether or not each unknown low-luminance region is a segregation region (S24a (classification step)).

  In this case, the analysis unit 107 operates the segregation detection unit 109, and similarly to S25 (FIG. 4), among the unknown low-luminance regions of the target of S24a, each classified as a segregation region in S24a. It detects as a segregation area | region (S25a (segregation detection step)).

  On the other hand, the classification unit 108 targets the unknown low-luminance region determined to be included in the region (second corresponding region) corresponding to the first V segregation analysis region and the second V segregation analysis region (second region) in S41. Then, S24b is performed. However, the classification unit 108 performs S42 before performing S24b when performing S24b on the unknown low-luminance region determined to be included in the region corresponding to the first V segregation analysis region in S41.

  Specifically, in S42, the classification unit 108 calculates the image feature amount of each unknown low-luminance area that is determined in S23 to be included in the area corresponding to the first V segregation analysis area in S41. In the same manner as in FIG. 5, the center position in the vertical direction in the unknown image is turned upside down (S 42).

  In S24b, as in S24 (FIG. 4), the classification unit 108 applies the V segregation classification model (second classification model) constructed in S15b (FIG. 5) to the image feature amount after performing S42. Is applied to classify whether each of the unknown low-luminance regions determined to be included in the region corresponding to the first V segregation analysis region in S41 is a segregation region. Similarly to S24 (FIG. 4), the classification unit 108 calculates the image features of the unknown low-luminance regions calculated in S23 and determined to be included in the region corresponding to the second V segregation analysis region in S41. By applying the V segregation classification model (second classification model) to the quantity, each unknown low-luminance area determined to be included in the area corresponding to the second V segregation analysis area in S41 is a segregation area. It is classified whether or not there is (S24b (classification step)).

  In this case, the analysis unit 107 operates the segregation detection unit 109, and similarly to S25 (FIG. 4), the region corresponding to the first V segregation analysis region and the second V segregation analysis region (second correspondence region) in S41. Of the unknown low-brightness areas determined to be included, the areas classified as segregation areas in S24b are detected as segregation areas (S25b (segregation detection step)).

  And the analysis part 107 will complete | finish an analysis process, if the detection of the segregation area | region from an unknown image is completed by complete | finishing S25a and S25b.

  According to the configuration of the second embodiment, the center segregation classification model is applied to the image feature amount of each unknown low-luminance region included in the region corresponding to the center segregation analysis region in the unknown image. Considering the image features specific to the low-brightness areas that are discriminated as segregation areas by visual inspection included in the analysis area, classify whether each unknown low-brightness area is a segregation area of central segregation with higher accuracy. be able to.

  Similarly, by applying the V segregation classification model to the image feature quantities of the unknown low luminance regions included in the regions corresponding to the first and second V segregation analysis regions in the unknown image, the first and second In consideration of the image feature amount peculiar to the low-luminance area visually identified as the segregation area included in the two V segregation analysis area, whether or not each of the unknown low-luminance areas is a segregation area of V segregation is determined. It is possible to classify with high accuracy.

(Third embodiment)
The third embodiment according to the present invention will be described below. In the configuration of the second embodiment, a plurality of areas included in each of the two areas of the center segregation analysis area and the first and second V segregation analysis areas different from the center segregation analysis area included in the teaching image. Based on the image feature amount of each low-luminance area, center segregation classification data and V segregation classification data corresponding to each of the two areas are created.

  Then, the center segregation classification model and the V segregation classification model corresponding to each of the two regions are individually constructed using the created center segregation classification data and V segregation classification data. Then, using the constructed center segregation classification model and V segregation classification model, respectively, a plurality of unknown low-luminance areas identified in two areas corresponding to the two areas in the unknown image, respectively. Was classified as a segregation region, and the segregation region was detected.

  However, in the configuration of the third embodiment, a region different from the center segregation analysis region included in the teaching image is divided into a first V segregation analysis region and a second V segregation analysis region, A center corresponding to each of the three regions based on the image feature amount of each of the plurality of low luminance regions included in each of the three regions of the V segregation analysis region and the second V segregation analysis region. Segregation classification data, first V segregation classification data, and second V segregation classification data are created.

  Then, using the created three classification data, three classification models corresponding to the three areas are individually constructed. Further, using the constructed three classification models, whether or not each of the plurality of unknown low-luminance regions identified in each of the three regions corresponding to the three regions in the unknown image is a segregation region. The segregation region is detected.

  Specifically, in the configuration of the third embodiment, the analysis unit 106 performs analysis processing according to the flow shown in FIG. FIG. 7 is a flowchart illustrating another example of the analysis process. As shown in FIG. 7, when the analysis process is started by the analysis unit 106, S11 to S13 are performed as in the first and second embodiments. Moreover, S31, S14a, and S15a are performed similarly to 2nd embodiment.

  On the other hand, unlike the second embodiment, the classification data creation unit 104 performs S14c on the low-luminance region determined to be included in the first V segregation analysis region (third region) in S31, and the second in S31. S14d is performed for the low luminance region determined to be included in the V segregation analysis region (fourth region).

  In S14c, the classification data creation unit 104 determines whether or not each low-luminance region determined to be included in the first V segregation analysis region in S31 is a segregation region in the same manner as in S14a. First V segregation classification data (third classification data) in which the result is associated with the image feature amount of each of the low luminance areas calculated in S13 is created (S14c (classification data creation step)).

  In this case, the analysis unit 106 operates the classification model construction unit 105 and uses the image feature amount included in the first V segregation classification data created in S14c as an explanatory variable, similarly to S15a, and uses the first V segregation. The first V segregation classification model (third classification model) is constructed using the discrimination result included in the classification data as the objective variable (S15c (classification model construction step)).

  Similarly, in S14d, the classification data creation unit 104 visually discriminates whether or not each of the low luminance areas determined to be included in the second V segregation analysis area is a segregation area in S31. And the second V segregation classification data (fourth classification data) in which the image feature values of the low luminance areas calculated in S13 are associated with each other (S14d (classification data creation step)).

  In this case, the analysis unit 106 operates the classification model construction unit 105 and uses the image feature amount included in the second V segregation classification data created in S14d as an explanatory variable, and is included in the second V segregation classification data. The second V segregation classification model (fourth classification model) is constructed using the determined discrimination result as the objective variable (S15d (classification model construction step)).

  And the analysis part 106 will complete | finish an analysis process, if the construction of the center segregation classification model in S15a, the construction of the first V segregation classification model in S15c, and the construction of the second V segregation classification model in S15d are completed.

  On the other hand, in the configuration of the third embodiment, the analysis unit 107 executes analysis processing according to the flow shown in FIG. FIG. 8 is a flowchart illustrating another example of the analysis process.

  As shown in FIG. 8, when the analysis process is started by the analysis unit 107, S21 to S23 are performed as in the first and second embodiments. Moreover, S41, S24a, and S25a are performed similarly to 2nd embodiment.

  On the other hand, unlike the second embodiment, the classification unit 108 targets an unknown low-luminance region that is determined to be included in the region (third corresponding region) corresponding to the first V segregation analysis region (third region) in S41. S24c is performed, and S24d is performed on the low-luminance region determined to be included in the region (fourth corresponding region) corresponding to the second V segregation analysis region (fourth region) in S41.

  In S24c, similarly to S24a, the classification unit 108 uses the first V segregation constructed in S15c (FIG. 7) for the image feature values of the unknown low-luminance regions of S24c that are calculated in S23. By applying the classification model (third classification model), it is classified whether or not each of the unknown low luminance areas is a segregation area (S24c (classification step)).

  In this case, the analysis unit 107 operates the segregation detection unit 109, and similarly to S25a (FIG. 6), among the unknown low-luminance regions of S24c, each classified as a segregation region in S24c. It detects as a segregation area | region (S25c (segregation detection step)).

  Similarly, in S24d, the classifying unit 108 calculates the second V segregation classification constructed in S15d (FIG. 7) for the image feature amounts of the unknown low-luminance regions targeted in S24d calculated in S23. By applying the model (fourth classification model), it is classified whether or not each of the unknown low luminance areas is a segregation area (S24d (classification step)).

  In this case, the analysis unit 107 operates the segregation detection unit 109 to detect, as the segregation region, each of the unknown low-brightness regions of S24d that are classified as segregation regions in S24d (S25d ( Segregation detection step)).

  And the analysis part 107 will complete | finish an analysis process, if detection of the segregation area | region from an unknown image is completed by complete | finishing S25a, S25c, and S25d.

  According to the configuration of the third embodiment, the low brightness determined as the segregation region of the V segregation (hereinafter referred to as the first V segregation) generated in the first V segregation analysis region by visual inspection, which is included in the first V segregation analysis region. In consideration of the image characteristic amount peculiar to the region, it is more accurate whether or not each of the unknown low luminance regions included in the region corresponding to the first V segregation analysis region in the unknown image is the segregation region of the first V segregation. Can be classified well.

  Similarly, the image characteristic amount peculiar to the low-luminance region, which is included in the second V segregation analysis region and determined to be a segregation region of V segregation (hereinafter, second V segregation) generated in the second V segregation analysis region by visual inspection In consideration, it is possible to classify more accurately whether each unknown low-luminance region included in the region corresponding to the second V segregation analysis region in the unknown image is a segregation region of the second V segregation.

  In addition, said 1st thru | or 3rd embodiment is only the illustration of embodiment which concerns on this invention, and is not the meaning which limits this invention to said 1st thru | or 3rd embodiment. For example, the following modified embodiment may be used.

  (1) In the configurations of the second and third embodiments, S31 (FIGS. 5 and 7) may be performed immediately after S12. And like S13, after calculating the image feature-value of each low-luminance area | region judged to be contained in the center segregation analysis area | region by the said S31, you may make it perform the process after S14a. Similarly to S13, after calculating the image feature amounts of the low luminance areas determined to be included in the first and second V segregation analysis areas in S31, in the configuration of the second embodiment, S32 (FIG. 5) and subsequent processes may be performed, and in the configuration of the third embodiment, the processes after S14c (FIG. 7) or the processes after S14d (FIG. 7) may be performed.

  (2) In the configurations of the second embodiment and the third embodiment, S41 (FIGS. 6 and 8) may be performed immediately after S22. Then, similarly to S23, after calculating the image feature amount of each unknown low-brightness area determined to be included in the area corresponding to the central segregation analysis area in S41, the processing after S24a is performed. May be. Similarly to S23, after calculating the image feature amounts of the unknown low-luminance regions determined to be included in the regions corresponding to the first and second V segregation analysis regions in S41, the second embodiment In the configuration of FIG. 8, the processing after S42 (FIG. 6) is performed, and in the configuration of the third embodiment, the processing after S24c (FIG. 8) or the processing after S24d (FIG. 8) is performed. Good.

  (3) The cross section of the steel material is not limited to a cross section when a plate-shaped slab produced by continuous casting is cut in parallel to the casting direction, but may be a cross section of a steel material manufactured by another method. In this case, in the configurations of the second and third embodiments, instead of the center segregation analysis region, a region where segregation peculiar to the cross section of the steel material occurs (hereinafter referred to as a specific segregation region) may be employed. Further, in the configuration of the third embodiment, instead of the first and second V segregation analysis regions, two different regions in which specific segregation different from the above-mentioned specific segregation occurs in the cross section of the steel material are adopted. May be.

  (4) S11 (FIGS. 3, 5, and 7) in the configuration of the first to third embodiments and the modified embodiments of (1) to (3) above is prepared by preparing a plurality of etched sections of steel cross sections. In this case, a plurality of teaching images representing each of the plurality of transfer bodies may be acquired. Then, in S12 (FIGS. 3, 5, and 7), after specifying all the low-luminance areas included in the acquired plurality of teaching images, the processing after S12 may be performed. In this case, since the amount of data included in the classification data increases, the segregation region can be detected with higher accuracy using a classification model constructed using the classification data.

(Other modified embodiments)
Hereinafter, other modified embodiments according to the present invention will be described. In the configurations of the first to third embodiments and the modified embodiments (1) to (4) described above, the segregation region is detected with high accuracy from the transfer image representing the transfer body of the steel material cross section. In the configuration of another modified embodiment according to the present invention, the quality of the steel material corresponding to the transfer image is further evaluated based on the image feature amount of each segregation area detected in the above configuration.

(Reason for recalling the configuration of another modified embodiment)
Hereinafter, the reason which recalled the structure of the said other modified embodiment is demonstrated. Conventionally, as a method for evaluating the quality of a steel material based on the state of segregation included in a steel material cross section, in addition to the method described in Patent Document 1 described above, JP-A-2014-172074 (hereinafter referred to as Document 1), Methods described in Japanese Unexamined Patent Application Publication No. 2013-145221 (hereinafter referred to as Document 2) and Japanese Unexamined Patent Application Publication No. 2013-250113 (hereinafter referred to as Document 3) are known.

  According to Document 1, when the maximum particle size and density of segregation contained in the cut surface of slab cutting are within a predetermined range expressed using a linear function, the slab is applied to sour-resistant steel. It is described that it is determined whether or not hydrogen induced cracking (HIC) occurs. In Document 2, the concentration of a predetermined element (C, Mn, Nb, P, etc.) contained in the steel material is measured by EPMA, and the result obtained by substituting the measured concentration into a predetermined judgment formula, Of HIC cracking is evaluated. In Document 3, whether or not the chemical composition contained in the steel material, the hardness and microstructure of the center segregation, and the actual conditions in the ultrasonic flaw detection and rolling / cooling processes near the surface layer of the steel material are within the predetermined range. Is used to determine the HIC resistance performance of a steel material.

  However, the method described in Document 1 assumes that the relationship between the maximum particle size and density of segregation and the quality of the slab is a linear relationship, and the relationship between the maximum particle size and density of segregation and the quality of the slab. There is a problem that it is impossible to judge with high accuracy when is a non-linear relationship. The methods described in Documents 2 and 3 have a problem that it takes time to analyze chemical components contained in the steel material, and it takes a long time to obtain a result of determining the quality of the steel material. In order to solve these problems, the inventor examined a configuration capable of accurately and quickly determining the quality of the steel material corresponding to the transfer image based on the image feature amount of the segregation area included in the transfer image. As a result, the inventor recalled the configuration of the other modified embodiment.

(First modified embodiment)
Hereinafter, the first modified embodiment among the other modified embodiments will be described. FIG. 9 is a block diagram illustrating another example of the configuration of the control unit 10. In the configuration of the first modified embodiment, when the CPU executes the control program stored in the ROM, the control unit 10 further includes a derivation unit 110 and a determination data creation unit 111 as shown in FIG. , Function as a determination model construction unit 112, a quality analysis unit 113, and a quality determination unit 114.

  The derivation unit 110 derives a feature amount related to the segregation region of the unknown image based on the image feature amount of each segregation region detected by the segregation detection unit 109 after the analysis processing by the analysis unit 107. The image feature amount of each segregation region is calculated by the image feature amount calculation unit 103 in the analysis process, and includes, for example, the area of each segregation region and the average value of luminance values. The feature values related to the segregation region of the unknown image include, for example, the average value of the segregation region included in the unknown image (hereinafter, unknown segregation region), the sum of the areas of the unknown segregation region, the area of the largest unknown segregation region, the unknown The number of segregation areas and the ratio of the average value of luminance values of the unknown segregation area and the average value of luminance values of areas other than the unknown segregation area in the unknown image are included.

  The determination data creation unit 111 determines whether the quality of the steel material (hereinafter, unknown steel material) corresponding to the unknown image is good by visual observation of the unknown image, and the unknown image derived by the derivation unit 110. Judgment data in which feature quantities related to segregation regions are associated is created. In addition, the steel material corresponding to an unknown image is the steel material used for preparation of the transfer body corresponding to an unknown image.

  For example, the determination data creation unit 111 causes the display unit 60 to display the unknown image acquired by the image acquisition unit 101 in S21 (FIG. 4) in the analysis process. The user visually checks the unknown image displayed on the display unit 60 and determines whether or not the quality of the unknown steel material is good. The user inputs the determined result using the input unit 50. The determination data creation unit 111 creates determination data in which the input determination result is associated with the feature amount related to the segregation region of the unknown image derived by the derivation unit 110.

  The determination model construction unit 112 uses the feature amount included in the determination data created by the determination data generation unit 111 as an explanatory variable, uses the determination result included in the determination data as an objective variable, and classifies the objective variable The model of the explanatory variable for this purpose (hereinafter referred to as a determination model) is constructed using a known statistical method.

  The statistical method includes, for example, explanatory variables (features related to segregation regions of unknown images) so that the purity of the objective variable (determination result obtained by visual determination of whether the quality of the unknown steel is good) increases. General analysis such as decision tree analysis that builds rules (judgment models) that combine tree and their thresholds into a tree shape, and logistic regression analysis to obtain regression equations (judgment models) to logistic curves that include objective variables and explanatory variables Techniques used in machine learning can be applied. For the statistical method, for example, a method used in general machine learning such as nearest neighbor method, neural network, support vector machine, and random forest can be applied. Alternatively, the determination model construction unit 112 may construct a judgment model using a plurality of methods among the above statistical methods.

  After the analysis processing by the analysis unit 107, the quality analysis unit 113 performs a quality analysis process (quality analysis step) in which the derivation unit 110, the determination data creation unit 111, and the determination model construction unit 112 are operated in order.

  FIG. 10 is a flowchart illustrating an example of the quality analysis process. Specifically, the quality analysis unit 113 starts the quality analysis process when the analysis process (FIGS. 4, 6, and 8) by the analysis unit 107 ends. As shown in FIG. 10, when starting the quality analysis process, the quality analysis unit 113 first operates the derivation unit 110 to detect in S25 (FIG. 4) and 25a to d (FIGS. 6 and 8) in the analysis process. Based on the image feature amount of each segregated region, the feature amount related to the segregated region of the unknown image is derived (S51 (derivation step)). In S51, the deriving unit 110 uses the image feature values of the segregation areas calculated in the processing of S23 (FIGS. 4, 6, and 8) in the analysis process, and the feature quantities related to the segregation area of the unknown image. Is derived.

  Next, the quality analysis unit 113 operates the determination data creation unit 111 to determine whether or not the quality of the unknown steel material is good by visual observation of the unknown image, and the unknown image derived in S51. Determination data in which the feature amount related to the segregation area is associated is generated (S52 (determination data generation step)).

  Then, the quality analysis unit 113 operates the determination model construction unit 112 to use the feature amount included in the determination data created in S52 as an explanatory variable and the determination result included in the determination data as an objective variable. Then, a determination model that is a model of the explanatory variable for classifying the objective variable is constructed (S53 (determination model construction step)), and the quality analysis process is terminated.

  Returning to FIG. After the quality analysis process by the quality analysis unit 113 is completed, the quality determination unit 114 executes the steel material quality determination process using the evaluation target transfer body. Here, the transfer body for evaluation is a transfer body that is different from the transfer body that has transferred the cross section of the steel material to be analyzed, which is used when the analysis unit 107 operates the image acquisition unit 101 in the analysis process. Specifically, the transfer object for evaluation is created by transferring the cross section of the steel material to be evaluated to the thin film after etching the cross section of the steel material to be evaluated, as in the case of the transfer body having transferred the cross section of the steel material to be analyzed.

  The quality determination unit 114 operates the analysis unit 107 and the derivation unit 110 in order in the steel material quality determination process, and then relates to the segregation region of the unevaluated image representing the evaluation object transfer body derived by the derivation unit 110. By applying the judgment model constructed by the judgment model construction unit 112 to the quantity, it is judged whether or not the quality of the unevaluated steel material corresponding to the unevaluated image is good. Note that the unevaluated steel material corresponding to the unevaluated image is a steel material used for creating a transfer body corresponding to the unevaluated image.

  FIG. 11 is a flowchart illustrating an example of the steel material quality determination process. Specifically, after the quality analysis process by the quality analysis unit 113 is completed, when an instruction to start the steel material quality determination process is input by the input unit 50, the quality determination unit 114 starts the steel material quality determination process. As shown in FIG. 11, when the quality determination unit 114 starts the steel quality determination process, first, the analysis unit 107 is operated to execute an analysis process (analysis step) using a transfer body for evaluation (S61). ).

  In S61, first, S21 (FIGS. 4, 6, and 8) is executed, and an unevaluated image representing a transfer body for evaluation is acquired by the image acquisition unit 101. Next, S22 (FIGS. 4, 6, and 8) is executed, and the image processing unit 102 identifies a plurality of low-brightness regions (hereinafter referred to as unevaluated low-brightness regions) having a luminance lower than a predetermined luminance in the unevaluated image. Is done. Next, S23 (FIGS. 4, 6, and 8) is executed, and the image feature amount calculation unit 103 calculates the image feature amount of each of the plurality of unevaluated low-luminance regions specified in the unevaluated image. . Thereafter, S24 and S25 (FIG. 4), the process after S41 shown in FIG. 6, or the process after S41 shown in FIG. 8 is executed. As a result, a segregation area included in the unevaluated image is detected.

  When the analysis process of S61 is completed, the quality determination unit 114 operates the derivation unit 110, and as in S51 (FIG. 10), the segregation of the unevaluated image is performed based on the image feature values of the segregation areas detected in S61. A feature amount relating to the region is derived (S62 (derivation step)).

  Next, the quality determination unit 114 responds to the unevaluated image by applying the determination model constructed in S53 (FIG. 10) to the feature amount related to the segregation region of the unevaluated image derived in S62. It is determined whether or not the quality of the unevaluated steel material is good (S63 (quality determination step)), and the steel material quality determination process is terminated.

  For example, it is assumed that a determination model is constructed by a decision tree analysis method in S53 (FIG. 10). In other words, the explanatory variable (feature amount related to the segregation region of the unknown image) and its threshold value are set so that the determination model increases the purity of the objective variable (determination result obtained by visually determining whether or not the quality of the unknown steel is good). Is a rule that is combined in a tree form. In this case, in S63, the quality determination unit 114 is determined in the determination model among the characteristic amounts related to the segregation region of the unevaluated image derived in S62 and the characteristic amount included as the explanatory variable in the determination model. It is determined (classified) whether or not the quality of the unevaluated steel material is good based on the comparison result with the threshold value of the feature amount.

  Alternatively, it is assumed that a determination model is constructed by a logistic regression analysis method in S53 (FIG. 10). That is, the judgment model includes objective variables (judgment results obtained by visually judging whether or not the quality of the unknown steel material is good) and explanatory variables (features related to the segregation area of the unknown image) constructed by logistic regression analysis. Let it be a regression equation to a logistic curve. In this case, in S63, the quality determination unit 114 uses, as the regression equation, the feature amount included as the explanatory variable in the regression equation indicated by the determination model among the feature amounts related to the segregated region of the unevaluated image derived in S62. Whether or not the quality of the unevaluated steel material is good is classified (determined) according to the result assigned to.

  Therefore, according to the configuration of the first modified embodiment, it is possible to use the feature amount derived in S62 without using time and effort to grasp the feature amount related to the segregated region of the unevaluated image by visual inspection. Whether or not the quality is good can be quickly determined. Further, not only the feature quantity related to the segregation area of the unevaluated image is considered, but also the segregation area of the transfer image corresponding to the steel material whose quality is determined to be good in the past using the determination model. It is possible to determine whether or not the quality of the unevaluated steel material is good with higher accuracy by further considering whether or not the feature amount is related to.

(Second modified embodiment)
Next, a second modified embodiment among the other modified embodiments will be described. In the configuration of the second modified embodiment, unlike the configuration of the first modified embodiment, each of the segregation regions detected from the two regions of the center segregation analysis region and the other regions in the unknown image, respectively. Based on the image feature amount, the quality of each of the two regions corresponding to the two regions in the unevaluated steel material is individually determined.

  In the configuration of the second modified embodiment, the quality analysis unit 113 executes quality analysis processing according to the flow shown in FIG. 12, unlike the first modified embodiment. FIG. 12 is a flowchart illustrating another example of the quality analysis process.

  Specifically, when the analysis process (FIGS. 4, 6, and 8) by the analysis unit 107 is completed, the quality analysis unit 113 starts the quality analysis process illustrated in FIG. When the quality analysis unit 113 starts the quality analysis process, the quality analysis unit 113 operates the derivation unit 110 to execute S50, S51a, and S51b.

  In S50, the derivation unit 110 determines that each segregation area is unknown based on the image feature amount of each segregation area detected in S25 (FIG. 4) and 25a to 25d (FIGS. 6 and 8) in the analysis process. In the image, a region corresponding to the center segregation analysis region (first unknown region), a region corresponding to the first V segregation analysis region (second unknown region), and a region corresponding to the second V segregation analysis region (second unknown region) ) Is determined (S50).

  Hereinafter, for convenience of explanation, a region corresponding to the center segregation analysis region in the unknown image is abbreviated as a center segregation analysis region in the unknown image. Similarly, the region corresponding to the first V segregation analysis region in the unknown image is abbreviated as the first V segregation analysis region in the unknown image, and the region corresponding to the second V segregation analysis region in the unknown image is the first in the unknown image. Abbreviated as 2V segregation analysis region.

  Specifically, in S50, the derivation unit 110 determines that the segregation areas detected in the analysis process are the center segregation analysis area, the first V segregation analysis area in the unknown image, and the same as S31 (FIG. 5). It is determined in which of the second V segregation analysis areas.

  Then, the deriving unit 110 performs S51a using the image feature amounts of the segregation areas determined to be included in the center segregation analysis area in the unknown image in S50. Specifically, in S51a, the deriving unit 110 derives a feature amount related to the segregation area of the center segregation analysis area in the unknown image (S51a). The feature amount related to the segregation area of the center segregation analysis area in the unknown image includes, for example, the average value of the area of the segregation area (hereinafter, unknown center segregation area) included in the center segregation analysis area in the unknown image, the area of the unknown center segregation area Of the maximum unknown center segregation area, the number of unknown center segregation areas, the average value of the luminance values of the unknown center segregation area, and the brightness values of the areas other than the unknown center segregation area in the center segregation analysis area in the unknown image. The ratio with the average value is included.

  When S51a ends, the quality analysis unit 113 operates the determination data creation unit 111 to perform S52a (determination data creation step) as in S52 (FIG. 10). In S52a, the derivation unit 111 determines whether or not the quality of the first unknown steel material region corresponding to the central segregation analysis region in the unknown steel is good by visual observation of the center segregation analysis region in the unknown image, The center segregation determination data (first determination data) in which the feature amount related to the segregation area of the center segregation analysis area in the unknown image derived in S51a is associated is created (S52a (determination data creation step)). In addition, the 1st unknown steel material area | region corresponding to the center segregation analysis area | region in an unknown steel material shows the area | region extended in the horizontal direction in the center vicinity of the perpendicular direction in an unknown steel material.

  For example, the determination data creation unit 111 causes the display unit 60 to display the unknown image acquired by the image acquisition unit 101 in S21 (FIG. 4) in the analysis process in S52a. The user visually checks the center segregation analysis region in the unknown image displayed on the display unit 60 and determines whether or not the quality of the first unknown steel material region is good. The user inputs the determined result using the input unit 50. The determination data creation unit 111 associates the input determination result regarding the quality of the first unknown steel material region with the feature value related to the segregation region of the center segregation analysis region in the unknown image derived in S51a. Create judgment data.

  When S52a ends, the quality analysis unit 113 operates the determination model construction unit 112 and performs S53a (determination model construction step) as in S53 (FIG. 10). In S53a, the determination model construction unit 112 uses the feature amount included in the center segregation determination data created in S52a as an explanatory variable, and uses the determination result included in the center segregation determination data as an objective variable. A central segregation determination model (first determination model) that is a model of the explanatory variables for classifying the variables is constructed (S53a).

  In addition, the deriving unit 110 performs S51b using the image feature values of the segregation areas determined to be included in the first V segregation analysis area or the second V segregation analysis area in the unknown image in S50. Hereinafter, the first V segregation analysis region and the second V segregation analysis region in the unknown image are collectively referred to as a V segregation analysis region (second unknown region) in the unknown image. Specifically, in S51b, the derivation unit 110 derives a feature amount related to the segregation region of the V segregation analysis region in the unknown image (S51b). The feature amount related to the segregation area of the V segregation analysis area in the unknown image includes, for example, the average value of the area of the segregation area (hereinafter, unknown V segregation area) included in the V segregation analysis area in the unknown image, the area of the unknown V segregation area Of the maximum unknown V segregation area, the number of unknown V segregation areas, the average brightness value of the unknown V segregation area, and the brightness value of the area other than the unknown V segregation area of the V segregation analysis area in the unknown image. The ratio with the average value is included.

  The feature amount included in each of the feature amount related to the segregation region of the central segregation analysis region derived by the deriving unit 110 in S51a and the feature amount related to the segregation region of the V segregation analysis region derived in S51b are not the same. Also good. For example, the feature value related to the segregation region in the center segregation analysis region includes the average value of the area of the unknown center segregation region, but the feature value related to the segregation region in the V segregation analysis region includes the average value of the area of the unknown V segregation region. It may not be included.

  When S51b ends, the quality analysis unit 113 operates the determination data creation unit 111 to perform S52b (determination data creation step), as in S52a. In S52b, the determination data creation unit 111 determines whether or not the quality of the second unknown steel material region corresponding to the V segregation analysis region in the unknown steel is good by visual observation of the V segregation analysis region in the unknown image. And V segregation determination data (second determination data) that associates the feature amount related to the segregation area of the V segregation analysis area in the unknown image derived in S51b (S52b). In addition, the 2nd unknown steel material area | region corresponding to the V segregation analysis area | region in an unknown steel material shows the area | region different from the 1st unknown steel material area | region (area | region extended in the horizontal direction in the vicinity of the center of a perpendicular direction) in an unknown steel material.

  For example, the determination data creation unit 111 causes the display unit 60 to display the unknown image acquired by the image acquisition unit 101 in S21 (FIG. 4) in the analysis process in S52b. The user visually checks the V segregation analysis region in the unknown image displayed on the display unit 60, and determines whether or not the quality of the second unknown steel material region corresponding to the V segregation analysis region in the unknown steel material is good. The user inputs the determined result using the input unit 50. The determination data creation unit 111 associates the input determination result on the quality of the second unknown steel material region with the feature amount related to the segregation region of the V segregation analysis region in the unknown image derived in S51b. Create judgment data.

  When S52b ends, the quality analysis unit 113 operates the determination model construction unit 112 and performs S53b (determination model construction step), as in S53a. In S53b, the determination model construction unit 112 uses the feature amount included in the V segregation determination data created in S52b as an explanatory variable, and uses the determination result included in the V segregation determination data as an objective variable. A V segregation determination model (second determination model) that is a model of the explanatory variables for classifying the variables is constructed (S53b).

  When the construction of the center segregation judgment model and the V segregation judgment model by the judgment model construction unit 112 is finished, the quality analysis unit 113 finishes the quality analysis process.

  That is, in the second modified embodiment, the feature amount in the center segregation analysis region in the unknown image and the feature amount in the V segregation analysis region different from the center segregation analysis region in the unknown image are individually obtained by S51a and S51b. To be derived. Further, the center segregation determination data and the V segregation determination data are individually created by S52a and S52b, and the center segregation determination model and the V segregation determination model are individually constructed by S53a and S53b.

  Accordingly, in the second modified embodiment, the quality determining unit 114 executes the steel material quality determining process according to the flow shown in FIG. 13, unlike the first modified embodiment. FIG. 13 is a flowchart illustrating another example of the steel material quality determination process.

  Specifically, after the quality analysis process shown in FIG. 12 by the quality analysis unit 113 is completed, when a steel material quality determination process start instruction is input by the input unit 50, the quality determination unit 114 starts the steel quality determination process. To do. As shown in FIG. 13, when the quality determination unit 114 starts the steel material quality determination process, the analysis unit 107 is operated similarly to the first modified embodiment, and the analysis process (analysis using the transfer body for evaluation target) is performed. Step) is executed (S61).

  When the analysis process of S61 ends, the quality determination unit 114 operates the derivation unit 110 to execute S60, S62a, and S62b.

  In S60, similarly to S50 (FIG. 12), the deriving unit 110 determines that each segregated area is an unevaluated image based on the image feature amount of each segregated area detected in the unevaluated image detected in S61. The region corresponding to the center segregation analysis region (first unevaluated region), the region corresponding to the first V segregation analysis region (second unevaluated region), and the region corresponding to the second V segregation analysis region (second unevaluated region) Which of the evaluation areas is included is determined (S60).

  Hereinafter, for convenience of explanation, the region corresponding to the center segregation analysis region in the unevaluated image is abbreviated as the center segregation analysis region in the unevaluated image. Similarly, the region corresponding to the first V segregation analysis region in the unevaluated image is abbreviated as the first V segregation analysis region in the unevaluated image, and the region corresponding to the second V segregation analysis region in the unevaluated image is This is abbreviated as the second V segregation analysis region in the unevaluated image.

  Then, the deriving unit 110 performs S62a using the image feature values of the segregation areas determined to be included in the center segregation analysis area in the unevaluated image in S60. Specifically, in S62a, the derivation unit 110 derives a feature amount related to the segregation area of the central segregation analysis area in the unevaluated image, similarly to S51a (FIG. 12) (S62a (derivation step)).

  When S62a ends, the quality determination unit 114 is constructed in S53a (FIG. 12) with respect to the feature amount related to the segregation region of the central segregation analysis region in the unevaluated image derived in S62a, as in S63 (FIG. 11). By applying the center segregation determination model, it is determined whether or not the quality of the first steel material region corresponding to the center segregation analysis region in the unevaluated steel material is good (S63a (quality determination step)). In addition, the 1st steel material area | region corresponding to the center segregation analysis area | region in unrated steel materials shows the area | region extended in the horizontal direction in the vicinity of the center of the perpendicular direction in unrated steel materials. By S63a, it is possible to more accurately and quickly determine whether or not the quality of the first steel material region in the unevaluated steel material is good in consideration of the feature amount related to the segregation region of the central segregation analysis region in the unevaluated image. it can.

  Further, the derivation unit 110 uses the image feature amounts of the segregation areas determined in S60 to be included in the first V segregation analysis area in the unevaluated image and the second V segregation analysis area in the unevaluated image. I do. Hereinafter, the first V segregation analysis region and the second V segregation analysis region in the unevaluated image are collectively referred to as a V segregation analysis region (second unevaluated region) in the unevaluated image.

  Specifically, in S62b, the derivation unit 110 derives a feature amount related to the segregation area of the V segregation analysis area in the unevaluated image in the same manner as in S51b (FIG. 12) (S62b (derivation step)).

  When S62b ends, the quality determination unit 114 is constructed in S53b (FIG. 12) for the feature amount related to the segregation region of the V segregation analysis region in the unevaluated image derived in S62b, as in S63 (FIG. 11). By applying the V segregation determination model, it is determined whether or not the quality of the second steel material region corresponding to the V segregation analysis region in the unevaluated steel material is good (S63b (quality determination step)). In addition, the 2nd steel material area | region corresponding to the V segregation analysis area | region in unrated steel materials shows the area | region different from the 1st steel material area | region in unrated steel materials. By S63b, it is possible to more accurately and quickly determine whether or not the quality of the second steel material region in the unevaluated steel material is good in consideration of the feature amount related to the segregation region of the V segregation analysis region in the unevaluated image. it can.

  When S63a and S63b are completed, the quality determination unit 114 determines the determination result (hereinafter referred to as the first determination result) about the quality of the first steel material region in the unevaluated steel material in S63a and the second steel material region in the unevaluated steel material in S63b. Based on the quality determination result (hereinafter referred to as the second determination result), it is determined whether or not the quality of the unevaluated steel is good, and the determination result is displayed on the display unit 60 (S63c).

  For example, in S63c, when the first determination result or the second determination result indicates that the quality is not good, the quality determination unit 114 determines that the quality of the unevaluated steel material is not good. However, not limited to this, in S63c, when the quality determination unit 114 indicates that the first determination result or the second determination result is good, it may be determined that the quality of the unevaluated steel material is good. Good. Alternatively, in S63c, the quality determination unit 114 may display the first determination result and the second determination result on the display unit 60 as they are.

  Further, for example, in S53a and S53b (FIG. 12), it is assumed that the center segregation determination model and the V segregation determination model are constructed by the decision tree analysis method. In this case, in S63a, the quality determination unit 114 includes the feature amount included in the center segregation determination model as an explanatory variable among the feature amounts derived in S62a and the feature amount determined in the center segregation determination model. The difference between the threshold value and the threshold value may be calculated. Similarly, in S63b, the quality determination unit 114 includes the feature amount included as the explanatory variable in the V segregation determination model among the feature amounts derived in S62b, and the feature amount defined in the V segregation determination model. The difference between the threshold value and the threshold value may be calculated. In S63c, the quality determination unit 114 may determine whether or not the quality of the unevaluated steel material is good based on the calculated sum of the two differences.

  Alternatively, in S53a and S53b (FIG. 12), it is assumed that the center segregation determination model and the V segregation determination model are constructed by the logistic regression analysis method. In this case, the quality determination unit 114 substitutes, in S63a, the feature amount included as an explanatory variable in the regression equation indicated by the central segregation determination model among the feature amounts derived in S62a. You may make it output as a determination result. Similarly, in S63b, the quality determination unit 114 substitutes a feature amount included as an explanatory variable in the regression equation indicated by the V segregation determination model among the feature amounts derived in S62b. You may make it output as a determination result. In S63c, the quality determination unit 114 may determine whether or not the quality of the unevaluated steel material is good based on the sum of the two calculated determination results.

DESCRIPTION OF SYMBOLS 1 Segregation detection apparatus 101 Image acquisition part 102 Image processing part 103 Image feature-value calculation part 104 Classification data creation part 105 Classification model construction part 106 Analysis part 107 Analysis part 108 Classification part 109 Segregation detection part 110 Derivation part 111 Determination data creation part 112 Determination model construction unit 113 Quality analysis unit 114 Quality determination unit S11, S21 Image acquisition step S12, S22 Image processing step S13, S23 Image feature amount calculation step S14, S14a, S14b, S14c, S14d Classification data creation step S15, S15a, S15b , S15c, S15d Classification model construction step S24, S24a, S24b, S24c, S24d Classification step S25, S25a, S25b, S25c, S25d Segregation detection step S51, S62, S62a, S62 Deriving step S52, S52a, S52b decision data creation step S53, S53a, S53b decision model building step S61 analysis step S63, S63a, S63b quality determination step

Claims (11)

An image acquisition step of acquiring an image representing a transfer body of an etched steel cross section;
An image processing step of identifying a plurality of low-brightness areas having a lower brightness than a predetermined brightness in the image;
An image feature amount calculating step for calculating an image feature amount of each of the plurality of low luminance regions;
Create classification data in which a discrimination result obtained by visually discriminating whether or not each of the plurality of low luminance regions is a segregation region and the calculated image feature amount of each of the plurality of low luminance regions Classification data creation step;
A classification model that is a model of the explanatory variable for classifying the objective variable using the image feature amount included in the classification data as an explanatory variable, the discrimination result included in the classification data as an objective variable A classification model construction step for constructing
Analysis steps to perform in sequence,
After completion of the analysis step, the image acquisition step, the image processing step, and the image feature amount calculation step are sequentially performed using the transfer body for analysis different from the transfer body used in the analysis step. ,
By applying the classification model to the image feature quantities of a plurality of unknown low-luminance areas specified in the unknown image representing the transfer body for analysis, the plurality of unknown low-luminance areas are each A classification step for classifying whether the region is a segregation region;
Of the plurality of unknown low-luminance regions, each of the classified low-incidence regions classified by the classification step as a segregation region is detected as a segregation region, and
An analysis step for sequentially executing
A segregation detection method comprising: In the classification data creation step,
Of the plurality of low luminance areas, first classification data in which the determination result and the image feature amount are associated with each of the low luminance areas included in the predetermined first area of the image, and the plurality of low luminance areas The second classification data that associates the discrimination result with the image feature amount for each of the low luminance areas included in the second area different from the first area in the image among the luminance areas individually. make,
In the classification model construction step,
Build a first classification model using the first classification data, build a second classification model using the second classification data,
In the classification step,
The first classification model is applied to the image feature amount of each unknown low luminance region included in the first corresponding region corresponding to the first region in the unknown image among the plurality of unknown low luminance regions. And classifying whether or not each unknown low luminance region included in the first corresponding region is a segregation region,
The second classification model is applied to the image feature amount of each unknown low-luminance area included in the second corresponding area corresponding to the second area in the unknown image among the plurality of unknown low-luminance areas. The segregation detection method according to claim 1, wherein each of the unknown low-luminance regions included in the second corresponding region is classified as a segregation region. The steel material cross section is a cross section when the slab is cut parallel to the casting direction,
The segregation detection method according to claim 2, wherein the first region is a center segregation analysis region extending in a horizontal direction in the vicinity of a vertical center of the image. In the classification data creation step,
Among the plurality of low luminance regions, the determination result and the image feature amount for each of the low luminance regions included in the third region of the second region that is vertically above the central segregation analysis region. The third classification data associated with each other, and among the plurality of low luminance regions, the low luminance regions included in the fourth region in the second region in the vertical direction lower than the central segregation analysis region, respectively. Separately creating the fourth classification data in which the discrimination result and the image feature amount are associated,
In the classification model construction step,
Build a third classification model using the third classification data, build a fourth classification model using the fourth classification data,
In the classification step,
The third classification model is applied to image feature quantities of unknown low-luminance areas included in a third corresponding area corresponding to the third area in the unknown image among the plurality of unknown low-luminance areas. By classifying whether or not each unknown low-luminance region included in the third corresponding region is a segregation region,
The fourth classification model is applied to image feature quantities of unknown low-luminance regions included in a fourth corresponding region corresponding to the fourth region in the unknown image among the plurality of unknown low-luminance regions. The segregation detection method according to claim 3, further comprising: classifying whether or not each unknown low-luminance region included in the fourth corresponding region is a segregation region. In the image processing step, binarization processing is performed on the image, and low-side luminance regions in the image after the binarization processing are specified as the plurality of low-luminance regions. The segregation detection method according to any one of claims 1 to 4. In the classification model construction step, a classification model is constructed using one or a plurality of statistical methods of decision tree, logistic regression, nearest neighbor method, neural network, support vector machine, and random forest. Item 6. The segregation detection method according to any one of Items 1 to 5. A derivation step for deriving a feature amount relating to the segregation region of the unknown image based on the image feature amount of each segregation region detected in the segregation detection step after completion of the analysis step;
A determination data creation step of creating determination data in which the determination result that determines whether or not the quality of the unknown steel material corresponding to the unknown image is good by visual observation of the unknown image is associated with the feature amount of the unknown image When,
A determination model that is a model of the explanatory variable for classifying the objective variable, with the feature amount included in the determination data as an explanatory variable, the determination result included in the determination data as an objective variable, A decision model construction step to construct;
A quality analysis step that executes
After completion of the quality analysis step, the analysis step and the derivation step are sequentially performed using the transfer body for evaluation that is different from the transfer body for analysis,
Whether the quality of the unevaluated steel corresponding to the unevaluated image is good by applying the determination model to the feature quantity of the unevaluated image representing the transfer body for the evaluation object. A quality judgment step for judging;
The segregation detection method according to any one of claims 1 to 6, further comprising: In the derivation step,
Individually deriving the feature amount of a predetermined first unknown region in the unknown image and the feature amount of a second unknown region different from the first unknown region in the unknown image;
In the determination data creation step,
A determination result for determining whether or not the quality of the first unknown steel region corresponding to the first unknown region in the unknown steel material is good by visual observation of the first unknown region, and the feature amount of the first unknown region And a determination result for determining whether or not the quality of the second unknown steel region corresponding to the second unknown region in the unknown steel material is good by visual observation of the second unknown region, Separately creating the second determination data that associates the feature amount of the second unknown region,
In the determination model construction step,
Build a first decision model using the first decision data, build a second decision model using the second decision data,
In the quality determination step,
By applying the first determination model to the feature amount of the first unevaluated area corresponding to the first unknown area in the unevaluated image, the first unevaluated area in the unevaluated steel material is supported. Determine whether the quality of the first steel material area to be good,
By applying the second determination model to the feature amount of the second unevaluated area corresponding to the second unknown area in the unevaluated image, it corresponds to the second unevaluated area in the unevaluated steel material. It determines whether the quality of the 2nd steel material area | region to perform is favorable, The segregation detection method of Claim 7 characterized by the above-mentioned. The steel material cross section is a cross section when the slab is cut parallel to the casting direction,
The segregation detection method according to claim 8, wherein the first unknown area is a center segregation analysis area extending in a horizontal direction in the vicinity of a center in a vertical direction in the unknown image. In the determination model construction step, the determination model is constructed using one or a plurality of statistical methods of decision tree, logistic regression, nearest neighbor method, neural network, support vector machine, and random forest. Item 10. The segregation detection method according to any one of Items 7 to 9. An image acquisition unit for acquiring an image representing a transfer body of an etched steel cross section;
An image processing unit that identifies a plurality of low-brightness areas having a lower brightness than a predetermined brightness in the image;
An image feature amount calculating unit for calculating an image feature amount of each of the plurality of low luminance regions;
Create classification data in which a discrimination result obtained by visually discriminating whether or not each of the plurality of low luminance regions is a segregation region and the calculated image feature amount of each of the plurality of low luminance regions A classification data creation unit;
A classification model that is a model of the explanatory variable for classifying the objective variable using the image feature amount included in the classification data as an explanatory variable, the discrimination result included in the classification data as an objective variable A classification model building unit for building
An analysis unit that performs an analysis process that sequentially operates the image acquisition unit, the image processing unit, the image feature amount calculation unit, the classification data creation unit, and the classification model construction unit;
A classification section;
A segregation detector;
After completion of the analysis process by the analysis unit, the image acquisition unit, the image processing unit, and the image feature amount calculation unit are analyzed using the transfer body for analysis different from the transfer body used in the analysis process. An analysis unit that performs an analysis process for operating the classification unit and the segregation detection unit in order after operating in sequence;
With
In the analysis process,
The classification unit applies the classification model to the image feature quantities of the plurality of unknown low-luminance regions specified in the unknown image representing the transfer body for analysis, thereby the plurality of unknown low-level images. Classify whether each luminance region is a segregation region,
The segregation detection unit detects each of the plurality of unknown low-luminance regions classified as a segregation region by the classification unit as a segregation region.

Images