Method, system, computer program, and method for generating trained model for estimating milling yield of test grain

Abstract

To provide a method, system, computer program, and method for generating a trained model for estimating milling yield of a test grain.SOLUTION: A method for estimating a milling yield of a test grain of the present invention includes processes of: cutting a seed of the test grain, and treating the cut surface with alkali; obtaining information corresponding to an amount of ferulic acid in the endosperm region of the alkali-treated cut surface of the seed of the test grain; looking up a correlation derived by using multiple milling yield data of a control grain and values corresponding to an amount of ferulic acid measured in the endosperm region of the alkali-treated cut surface of the seed of the control grain, and estimating the milling yield of the test grain based on information corresponding to the amount of ferulic acid in the test grain.SELECTED DRAWING: None

Description

[Technical field]

[0001] The present invention relates to a method, a system, a computer program, and a method for generating a trained model for estimating milling yield of a test grain. [Background technology]

[0002] Wheat is milled to separate it into flour and bran. Flour is used in food processing, and actual users demand a high "milling yield," which indicates the amount of flour that can be obtained from a certain amount of seeds (high milling performance). Previous research has revealed that genetic and environmental factors are intricately involved in milling quality. Therefore, it has been thought that the most realistic and reliable way to select milling quality in wheat breeding is to actually mill the wheat and calculate the milling yield value. In current wheat breeding, selection of milling quality is performed by actually milling lines from F6 (mid-breeding generation) onwards, calculating the milling yield value and comparing them (Non-Patent Document 1).

[0003] On the other hand, recent research has revealed that there is a correlation between the content of arabinoxylan, a polysaccharide that constitutes the cell walls of wheat flour, and the milling yield (Non-Patent Documents 2 and 3). Furthermore, a method for evaluating the milling yield by milling a small amount of flour has also been reported (Non-Patent Document 4). [Prior art documents] [Patent documents]

[0004] [Non-Patent Document 1] National Institute of Agricultural Science (2013), "Wheat Quality Evaluation Technology", FY2013 Innovative Agricultural Technology Acquisition Training [Non-Patent Document 2] Kato et al. (2002), "Relationship between wheat flour yield and cell wall polysaccharide content", National Agriculture and Food Research Organization Research Results Information [Non-Patent Document 3] Otobe (Kiribuchi) et al. (2002), "Improvement of flour yield of waxy wheat by reducing arabinoxylan content", Breeding Research, 4 (supplement 1), 200 [Non-Patent Document 4] Hashizume et al. (2013), "Development of a 5g flour milling evaluation method for wheat breeding", Aichi Agricultural Research Institute Report, 45, 21-25 Summary of the Invention [Problem to be solved by the invention]

[0005] The techniques described in the above documents all assume that the milling yield is calculated using wheat that has been milled, or that the actual measured value of the arabinoxylan content in the wheat flour is used. Milling usually requires 1 kg of wheat seeds, and it is difficult to evaluate milling ability in early breeding generations that only have a few to several dozen grains, so the F6 line has been used to select milling ability in wheat breeding, as described above. In addition, wheat seeds must be milled to quantify arabinoxylan, and complicated processing such as enzyme reactions is also required, so the actual measured value of the arabinoxylan content in wheat flour is not very suitable as a selection index for high milling ability. Therefore, if there was a method for selecting lines with high milling yields from early breeding generations for grains such as wheat, it would be possible to further accelerate the development of varieties with high flour milling properties. [Means for solving the problem]

[0006] As a result of intensive research by the present inventors to solve the above problems, they found that the amount of ferulic acid liberated from the endosperm region of grain seeds such as wheat by alkali treatment correlates with the milling yield of the grain, and that it is possible to easily quantitatively evaluate the amount of ferulic acid in grain seeds. In other words, they found that it is suitable as an index of grain milling properties, and have completed the present invention. The present invention provides the following. [1] A method for estimating milling yield of a test grain, comprising: (i) cutting a seed of a test grain and treating the cut surface with alkali; (ii) measuring a value corresponding to the amount of ferulic acid in the endosperm region of the alkali-treated cut surface of the seed of the test grain; (iii) estimating the milling yield of the test grain from the value measured in step (ii) using a calibration curve calculated using a plurality of milling yield data of the control grain and values ​​corresponding to the amount of ferulic acid measured in the endosperm region of the alkali-treated cut surface of the seed of the control grain; The method includes: [2] The method according to [1], wherein the value corresponding to the amount of ferulic acid is based on the absorbance attributable to ferulic acid. [3] The method according to [1] or [2], wherein the volatile alkaline aqueous solution is 0.1 M ammonia water, and the alkali treatment step includes dropping 0.1 M ammonia water onto the cross section. [4] (1) the test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; The method according to any one of the above [1] to [3]. [5] A method for estimating milling yield of a test grain, implemented by a computer, comprising: Obtaining a multi-wavelength image by photographing an endosperm region of an alkali-treated cut surface of a seed of a test grain using a multispectral camera; Inputting the multi-wavelength image to a first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; Calculating a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; Estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined; Including, The first trained model is a model that has been subjected to machine learning processing using training data so that, when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multi-spectral camera is input, the first trained model outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image. [6] A method for estimating milling yield of a test grain, comprising: (i) acquiring a multi-wavelength image by photographing an endosperm region of an alkali-treated cut surface of a seed of a test grain using a multispectral camera; Inputting the multi-wavelength image to a first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; Calculating a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; and (ii) estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined; Including, The first trained model is a model that has been subjected to machine learning processing using training data so that, when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multi-spectral camera is input, the first trained model outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image. [7] The teacher data is First training data in which a first multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for first training data showing a first flour yield using a multispectral camera, and a first value is associated with the first training data; The second training data showing the second milling yield includes a second multi-wavelength image produced by photographing an endosperm region of an alkali-treated cut surface of a grain seed using a multispectral camera, and a second training data in which a second value is associated with the second multi-wavelength image produced by photographing an endosperm region of an alkali-treated cut surface of a grain seed using a multispectral camera, The estimated ferulic acid concentration related index distribution of the multi-wavelength image is data expressed as a value corresponding to a ferulic acid concentration estimated for each pixel of the multi-wavelength image, The method according to [5] or [6], wherein calculating the representative value is an average value obtained by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels. [8] Obtaining a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera, Obtaining a multi-wavelength image by photographing the alkali-treated cut surface of the seed of the test grain with a multispectral camera; Inputting the multi-wavelength image to a second trained model and outputting an endosperm region applicability distribution of the multi-wavelength image; Extracting a multi-wavelength image of the endosperm region of the cut surface of the seed of the test grain based on the endosperm region relevance distribution; The method according to any one of [5] to [7], wherein the second trained model is one that has been subjected to machine learning processing using training data so as to output an endosperm region relevance distribution of a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera when the multi-wavelength image is input. [9] The method further includes deriving a correlation between the representative value and the milling yield of the test grain; Deriving the correlation includes: Obtaining a multi-wavelength image of the endosperm region by photographing the endosperm region of the alkali-treated cut surface of the seeds of a plurality of control grains having known milling yields using a multispectral camera; The multi-wavelength image is input to the first trained model, and an estimated ferulic acid concentration related index distribution of the multi-wavelength image is output; Calculating a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; The calculated representative value and the milling yield of the control grain are used to create a calibration curve. The method according to any one of the above [5] to [8].

[10] The method according to any one of [5] to [9], wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

[11] (1) the test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; The method according to any one of the above [5] to

[10] .

[12] A system for estimating an estimated ferulic acid concentration-related index distribution in an endosperm region of a cut surface of a seed of a test grain that has been treated with alkali, comprising: An acquisition unit that acquires a multi-wavelength image of the endosperm region by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera; An output unit that inputs the multi-wavelength image to a first trained model and outputs an estimated ferulic acid concentration related index distribution of the multi-wavelength image; A calculation unit that calculates a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The first trained model is a system that has been subjected to machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain with a multi-spectral camera is input, the system outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

[13] A system for estimating flour milling yield of a test grain, comprising: An acquisition unit that acquires a multi-wavelength image by photographing an endosperm region of an alkali-treated cut surface of a seed of a test grain using a multispectral camera; An output unit that inputs the multi-wavelength image to a first trained model and outputs an estimated ferulic acid concentration related index distribution of the multi-wavelength image; A calculation unit that calculates a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The first trained model is a system that has been subjected to machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain with a multi-spectral camera is input, the system outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

[14] The teacher data is A first training data showing a first milling yield, in which a first multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed using a multispectral camera and a first value is associated with the first training data; and second teacher data in which a second multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed using a multispectral camera, the second teacher data indicating a second milling yield, and a second value is associated with the second teacher data, The estimated ferulic acid concentration related index distribution of the multi-wavelength image is data in which a value corresponding to an estimated ferulic acid concentration is expressed for each pixel of the multi-wavelength image, The system according to claim 12 or 13, wherein the calculation unit calculates an average value by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels as the representative value.

[15] The acquisition unit is A multi-wavelength image is obtained by photographing the alkali-treated cut surface of the seed of the test grain using a multispectral camera; The multi-wavelength image is input to a second trained model, and an endosperm region applicability distribution of the multi-wavelength image is output; Based on the endosperm region conformity distribution, a multi-wavelength image of the endosperm region of the cut surface of the seed of the test grain is extracted, wherein: The system described in any one of

[12] to

[14] , wherein the second trained model is one that has been subjected to machine learning processing using training data so as to output an endosperm region relevance distribution of a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera when the multi-wavelength image is input.

[16] The system described in any one of

[12] to

[15] , further comprising an estimation unit that estimates the milling yield of the test grain based on the calculated representative value and a correlation between the representative value obtained in advance and the milling yield of the test grain.

[17] Further comprising a derivation unit for deriving a correlation between the representative value and the milling yield of the test grain; The acquisition unit acquires multi-wavelength images of the endosperm regions of the seeds of a plurality of reference grains having known milling yields by photographing the endosperm regions of the alkali-treated cut surfaces of the seeds with a multispectral camera; The output unit inputs the multi-wavelength image to the first trained model and outputs an estimated ferulic acid concentration related index distribution of the multi-wavelength image, The calculation unit calculates a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The system according to any one of

[12] to

[16] , wherein the derivation unit creates a calibration curve by using multiple of the calculated representative values ​​and the milling yields of the control grain.

[18] The system described in any one of

[12] to

[17] , wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

[19] (1) the test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; The system described in any one of

[12] to

[18] .

[20] To the computer, A process of acquiring a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera; A process of inputting the multi-wavelength image to a first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; A process of calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; A computer program for executing The first trained model is a model that has been subjected to machine learning processing using teacher data so as to output an estimated ferulic acid concentration-related index distribution of a multi-wavelength image created by photographing an endosperm region of an alkali-treated cut surface of a seed of the test grain with a multispectral camera when the multi-wavelength image is input. Computer program.

[21] The teacher data is A first training data showing a first milling yield, in which a first multi-wavelength image of an endosperm region of an alkali-treated cut surface of a grain seed taken by a multispectral camera is associated with a first value; The second training data showing the second milling yield includes a second multi-wavelength image of an endosperm region of an alkali-treated cut surface of a grain seed taken by a multispectral camera and a second value associated with the second training data, The estimated ferulic acid concentration related index distribution of the multi-wavelength image is data in which a value corresponding to an estimated ferulic acid concentration is expressed for each pixel of the multi-wavelength image, The computer program according to claim 20, wherein the process of calculating the representative value is to calculate an average value obtained by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels.

[22] The acquiring process includes: A process of acquiring a multi-wavelength image by photographing the alkali-treated cut surface of the seed of the test grain using a multispectral camera; A process of inputting the multi-wavelength image to a second trained model and outputting an endosperm region applicability distribution of the multi-wavelength image; A process of extracting a multi-wavelength image of the endosperm region of a cut surface of a seed of the test grain based on the endosperm region conformity distribution; The computer program described in

[20] or

[21] , wherein the second trained model is one that has been subjected to machine learning processing using training data so as to output a distribution of endosperm region relevance of a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera when the multi-wavelength image is input.

[23] A computer program described in any one of

[20] to

[22] , further causing the computer to execute a process of estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value previously obtained and the milling yield of the test grain.

[24] The computer further executes a process of deriving a correlation between the representative value and the milling yield of the test grain; The deriving process includes: A process of acquiring a multi-wavelength image of the endosperm region by photographing the endosperm region of the alkali-treated cut surface of the seeds of a plurality of control grains having known milling yields using a multispectral camera; A process of inputting the multi-wavelength image to the first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; A process of calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; A process of creating a calibration curve using the calculated representative value and the milling yield of the control grain; The computer program according to any one of claims 20 to 23.

[25] An input layer to which a multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera is input; an output layer that outputs the estimated ferulic acid concentration related index distribution of the input multi-wavelength image; An intermediate layer in which parameters are learned using teacher data so as to input the multi-wavelength image and output the estimated ferulic acid concentration related index distribution of the input multi-wavelength image; A computer program that causes a computer to function to perform a calculation based on the parameters of the intermediate layer on the multi-wavelength image input to an input layer of a trained model comprising the above-mentioned, and to output an estimated ferulic acid concentration-related index distribution of the multi-wavelength image from the output layer.

[26] The computer program according to any one of

[20] to

[25] , wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

[27] (1) the test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The computer program according to any one of

[20] to

[26] , wherein the test grain is wheat.

[28] A method for generating a trained model for estimating an estimated ferulic acid concentration-related index distribution in an endosperm region of an alkali-treated cut surface of a seed of a test grain, comprising: Obtaining teacher data in which a multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for teacher data with a multispectral camera and an estimated ferulic acid concentration-related index distribution of the multi-wavelength image is associated with each other; When a multi-wavelength image is input, the multi-wavelength image is created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera, and the multi-wavelength image is input based on the teacher data. A trained model is generated that outputs an estimated ferulic acid concentration-related index distribution of the input multi-wavelength image. A method comprising:

[29] The method according to

[28] , wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

[30] (1) the test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The method according to

[28] or

[29] , wherein the test grain is wheat. Effect of the Invention

[0007] According to the present invention, the milling quality of a grain can be easily estimated using only a small amount of the seeds of the grain and without going through a milling process. If varieties with high milling properties are developed using a selection method that uses the milling yield estimation method of this invention, they will be actively used by actual users. Ultimately, it is expected that this will lead to an increase in the amount of domestically produced wheat products in circulation. In particular, the reverse mismatch situation in which demand exceeds supply for hard wheat, which is processed into bread and Chinese noodles, continues, so this will meet the needs of actual users to develop new varieties. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram showing the correlation between measured wheat flour yield and measured free ferulic acid concentration in wheat seeds. [Diagram 2] FIG. 11 is a diagram showing an overall configuration of an estimation system according to a second embodiment of the present invention. [Diagram 3] FIG. 11 is a diagram illustrating a hardware configuration of an estimation system according to a second embodiment of the present invention. [Figure 4] FIG. 11 is a schematic diagram for explaining the generation process of a first trained model. [Diagram 5] 10 is a flowchart showing a process for estimating flour milling yield of a test grain according to a first aspect of the second embodiment. [Figure 6] 13 is a flowchart showing a correlation data creation process according to the first aspect of the second embodiment. [Figure 7] 13 is a flowchart showing a process for estimating the milling yield of a test grain according to a second aspect of the second embodiment. [Figure 8] 13 is a flowchart showing a correlation data creation process according to the second aspect of the second embodiment. [Figure 9] 13 is a flowchart showing a process for generating a first trained model according to a third aspect of the second embodiment. [Figure 10] 13 is a flowchart showing a process for generating a second trained model according to a third aspect of the second embodiment. [Figure 11] 1 is a diagram showing a multi-wavelength image of a cut surface of a seed, which is a multi-wavelength image created by photographing an endosperm region and converted into an RGB scale image. [Figure 12A] Fig. 12 shows the process of extracting the endosperm region from the multi-wavelength image. In the left diagram of Fig. 12A, the multi-wavelength image is converted to the RGB scale. In the right diagram of Fig. 12A, the endosperm region of the multi-wavelength image is colored green, and the other regions are colored yellow. [Figure 12B] FIG. 12B is an image in which the endosperm region conformity distribution is converted to an RGB scale and displayed, in which the endosperm region and other regions are colored continuously from red to blue. [Figure 12C] FIG. 12C is an image in which the endosperm region has been extracted, with the endosperm region colored in white and the other regions colored in black. [Figure 13A] Fig. 13 is a diagram showing a process of creating the first multi-wavelength image and the second multi-wavelength image. Fig. 13A shows images obtained by converting the first multi-wavelength image of the grain for training data and the second multi-wavelength image of the grain for training data before extracting the endosperm region into RGB scale. [Figure 13B] FIG. 13B shows images obtained by converting the first and second multi-wavelength images created by extracting the endosperm region into RGB scale images. [Figure 14A] 14A is a diagram showing an estimated ferulic acid concentration-related index distribution created from the first multi-wavelength image and the second multi-wavelength image. In FIG. 14A, the endosperm region of the first multi-wavelength image is colored light blue, and the endosperm region of the second multi-wavelength image is colored dark red. [Figure 14B]FIG. 14B is a diagram showing the distribution of the estimated ferulic acid concentration-related indexes converted into an RGB scale, in which the endosperm region is colored continuously from dark red to dark blue. [Figure 15] 1 is a diagram showing the correlation between the representative value of a control grain and the milling yield value of the control grain, in which the dashed line is an approximation curve for a scatter diagram plotting the representative value of the control grain and the milling yield, and the approximation curve functions as a calibration curve. [Figure 16] 1 is a diagram showing the correlation between the representative value of a control grain and the actual measured value of the ferulic acid concentration in the seeds of the control grain, in which the dashed line is an approximation curve for a scatter diagram in which the representative value of the control grain and the ferulic acid concentration are plotted, and the approximation curve functions as a calibration curve. [Figure 17] 1 is a diagram showing the correlation between the representative value of a control grain and the measured value of the arabinoxylan concentration in the seeds of the control grain, in which the dashed line is an approximation curve for a scatter diagram in which the representative value of the control grain and the arabinoxylan concentration are plotted, and the approximation curve functions as a calibration curve. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] The present invention relates to a method for estimating the flour yield of a test grain, and more particularly to a method for predicting the flour yield of a test grain by acquiring information corresponding to the amount of ferulic acid contained in the seeds of the test grain and utilizing the information and a correlation between the flour yield of a control grain that has been previously determined and information corresponding to the amount of ferulic acid contained in the seeds of the control grain.

[0010] The term "cereal grain" as used herein is a general term for annual or biennial herbaceous crops cultivated for harvesting grains, and particularly refers to grains containing arabinoxylan in the endosperm region of the seeds. Such grains are preferably wheat. The term "cereal seeds" as used herein refers to grains harvested from the grains. The wheat used in the present invention is not particularly limited, and both wild and cultivated species can be used. For example, various types of wheat such as hard wheat, medium wheat, soft wheat, durum wheat, and spelt wheat can be used.

[0011] "Ferulic acid" as described herein is a derivative of cinnamic acid, an organic compound having the following chemical formula: [ka]

[0012] In cereals, ferulic acid exists as an ester bond to arabinoxylan, which constitutes the cell wall. Arabinoxylan is a type of hemicellulose and is the main matrix polysaccharide in the cell wall of cereals. When some of the ferulic acid bound to arabinoxylan dimerizes to form diferulic acid, the arabinoxylans are cross-linked by the diferulic acid to form an interpolysaccharide network. It is believed that this interpolysaccharide network strengthens the cell wall. Arabinoxylan is also contained in the endosperm region of cereal seeds, and ferulic acid is present in the endosperm region bound to arabinoxylan.

[0013] [Estimation of flour yield] As specifically shown in the examples below, it has been newly revealed that the content of ferulic acid in the seeds of grains, particularly in the endosperm region, correlates with the milling yield of the grain. FIG. 1 is a diagram illustrating the correlation between the measured value of the milling yield of wheat and the measured value of the concentration of free ferulic acid in the endosperm region of the wheat seed. It was already known that the milling yield of wheat and the arabinoxylan content in the wheat seed have a negative correlation (Kato et al. (2002)), but as shown in FIG. 1, it has been discovered that there is a negative correlation between the milling yield of wheat and the content of ferulic acid in the endosperm region of the wheat seed. This means that the amount of ferulic acid in the endosperm region of the grain seed can be an index for estimating the milling yield of the grain. Ferulic acid is also present in the hull of grains, but since the hull does not contribute to milling yield, the content of ferulic acid in the endosperm region can be an indicator of milling yield.

[0014] Ferulic acid is easily liberated by alkali treatment since it is ester-bonded to arabinoxylan. Ferulic acid bound to arabinoxylan can be liberated by dropping an aqueous alkali solution onto the endosperm region of the cut surface of a grain seed. Ferulic acid is also known to have an absorption spectrum around 320 nm. The absorbance measured by liberating ferulic acid in the endosperm region of the cut surface of a grain seed and irradiating it with light in a wavelength region capable of identifying an absorption spectrum around 320 nm (for example, light with a wavelength of 365 nm to 940 nm) and the multi-wavelength image obtained by photographing the endosperm region in the wavelength region can be said to be information corresponding to the amount of ferulic acid contained in the endosperm region of the grain seed. Based on this information, the amount of ferulic acid contained in the endosperm region of the seeds of the test grain can be evaluated, and the milling yield of the test grain can be estimated from the correlation between the amount of ferulic acid and the milling yield.

[0015] In other words, the present invention is characterized by estimating milling yield using ferulic acid liberated in the endosperm region of the cut surface of the seed of the test grain as an indicator, making it possible to estimate the milling yield without milling the test grain and actually measuring the ferulic acid content. Hereinafter, the present invention will be described more specifically based on the embodiments of the present invention.

[0016] (First embodiment) A first embodiment of the present invention will now be described. The method for estimating flour milling yield of a test grain in the first embodiment includes: (i) cutting a seed of a test grain and treating the cut surface with alkali; (ii) measuring a value corresponding to the amount of ferulic acid in the endosperm region of the alkali-treated cut surface of the seed of the test grain; (iii) estimating the milling yield of the test grain from the value measured in step (ii) using a calibration curve calculated using a plurality of milling yield data of the control grain and values ​​corresponding to the amount of ferulic acid measured in the endosperm region of the alkali-treated cut surface of the seed of the control grain; Includes.

[0017] In step (i), the seeds of the test grain are cut by any method using a knife, a cutter, or the like. At this time, it is desirable to cut the center part of the seed so that the endosperm region is included in the cut surface. An appropriate amount (e.g., about 2 to about 5 μl) of an alkaline aqueous solution is dropped onto the cut surface according to its area, and the solution is left to stand for an appropriate time (e.g., about 15 to about 90 minutes) to liberate ferulic acid contained in the cut surface of the endosperm region of the seeds of the test grain. The aqueous alkali solution used for liberating ferulic acid is not particularly limited, but is preferably a volatile aqueous alkali solution, more preferably ammonia water. The concentration of the aqueous alkali solution is not particularly limited as long as it is a concentration capable of liberating ferulic acid from seeds, but can be, for example, about 0.01M to about 1M.

[0018] In step (ii), the endosperm region of the cut surface of the seed of the test grain from which ferulic acid has been liberated in step (i) is irradiated with light having a wavelength capable of identifying the absorption characteristics of ferulic acid, and the absorbance is measured. Since ferulic acid has an absorption spectrum around 320 nm as described above, it is preferable to irradiate with light having a wavelength between 365 nm and 940 nm in order to identify the absorption spectrum. The absorbance can be measured according to a conventional method using a spectrophotometer or the like. The "value corresponding to the amount of ferulic acid" in step (ii) may be the absorbance measured in the endosperm region itself, or a specific value converted from the absorbance, or the concentration of ferulic acid in the endosperm region. The "value corresponding to the amount of ferulic acid" is not limited to these values, and may be any other appropriate value.

[0019] In step (iii), a calibration curve is first prepared using a plurality of values ​​corresponding to the milling yield of a control grain and the amount of ferulic acid measured in the endosperm region of the control grain. The control grain can be the same grain as the test grain, and can be a grain whose milling yield and the value corresponding to the amount of ferulic acid in the cut surface of the seed are known. The control grains may be two or more grains having different values ​​corresponding to the milling yield and the amount of ferulic acid. From the viewpoint of the accuracy of the derived calibration curve, it is preferable to use a large number of control grains, for example, 5 or more, preferably 10 or more, more preferably 20 or more. Using the prepared calibration curve, the milling yield of the test grain is estimated from the value corresponding to the amount of ferulic acid in the test grain measured in step (ii).

[0020] Second embodiment A second embodiment of the present invention will now be described. The method for estimating flour milling yield of a test grain in the second embodiment includes at least: Obtaining a multi-wavelength image by photographing an endosperm region of an alkali-treated cut surface of a seed of a test grain using a multispectral camera; Inputting the multi-wavelength image to a first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; Calculating a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; Estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined; Includes.

[0021] <First aspect> The first aspect of the second embodiment will now be described. 1. Overview of the estimation system 2 is a diagram showing an overall configuration of an estimation system 1 according to this embodiment. The estimation system 1 according to this embodiment is communicatively connected to an image acquisition device 2 via a network N such as an intranet or the Internet. In this embodiment, the estimation system 1 and the image acquisition device 2 are separate devices, but the present invention is not limited to this. The estimation system 1 may be a system integrated with the image acquisition device 2, that is, a system including a multispectral camera, which will be described later.

[0022] The image acquisition device 2 is a multispectral camera. A multispectral camera is a camera that captures an image of a subject by irradiating the subject with light dispersed into a plurality of wavelength bands and detecting the reflected light. The multispectral camera has a plurality of narrowband filters that selectively transmit light in a plurality of narrowbands including the emission bands of a blue LED, a yellow LED, and a red LED, and simultaneously captures a plurality of images that have passed through the plurality of narrowband filters, respectively. The image acquisition device 2 does not need to be configured as one physical device, and may be configured as a plurality of physical devices.

[0023] In this specification, the "multi-wavelength image" is an image acquired by the multi-spectral camera, which is created by irradiating a subject with light dispersed into multiple wavelength bands and recording the reflected light. In general, spectral information that represents the reflection and absorption characteristics specific to the subject can be extracted from a multi-wavelength image, and the object captured in the multi-wavelength image can be recognized with high accuracy based on the spectral information of the object.

[0024] The image acquisition device 2 captures a multi-wavelength image by irradiating the alkali-treated cut surface of the seed of the test grain with multiple irradiation lights in wavelength bands capable of obtaining spectral information representing the reflection and absorption characteristics of ferulic acid. As described above, since ferulic acid has an absorption spectrum around 320 nm, it is preferable to use multiple irradiation lights with wavelengths between 365 nm and 940 nm, for example, in order to identify the absorption spectrum. The multi-wavelength image thus created contains spectral information representing the reflection and absorption characteristics of ferulic acid liberated from the endosperm region by the alkali treatment.

[0025] 2. Specific configuration of the estimation system The estimation system 1 executes a method for estimating the flour milling yield of a test grain by utilizing a multi-wavelength image of an endosperm region of a cut surface of an alkali-treated seed of the test grain, captured by an image capture device 2. As shown in FIG. 2, the estimation system 1 includes an acquisition unit 101, an output unit 102, a calculation unit 103, an estimation unit 104, a derivation unit 105, a model generation unit 106, and a storage unit 107. The estimation system 1 does not need to be configured as one physical device, and may be configured as multiple physical devices. The estimation system 1 may be a multi-computer configured to include multiple computers, or may be a virtual machine virtually constructed by software.

[0026] The acquisition unit 101 acquires a multi-wavelength image of the endosperm region by communicating with the image acquisition device 2 via a communication unit 10g described later. The acquisition unit 101 also acquires a multi-wavelength image captured by the image acquisition device 2, inputs the multi-wavelength image to a second trained model 107c that has been trained in advance by machine learning, outputs an endosperm region applicability distribution, and performs processing to acquire a multi-wavelength image of the endosperm region based on the endosperm region applicability distribution. The output unit 102 inputs the multi-wavelength image of the endosperm region to the first trained model 107b that has been trained in advance by machine learning, and performs processing to output an estimated ferulic acid concentration related index distribution. The calculation unit 103 performs a process of calculating a representative value based on the estimated ferulic acid concentration related index distribution. The estimation unit 104 performs a process of estimating the milling yield of the test grain based on the correlation between the representative value and the milling yield of the control grain. The derivation unit 105 derives the correlation between the representative value and the milling yield of the control grain. The model generation unit 106 performs learning using previously prepared teacher data, and performs processing to generate a first trained model 107b and a second trained model 107c.

[0027] The storage unit 107 stores various information such as various programs executed by the estimation system 1. The storage unit 107 stores a flour milling yield estimation program 107a, two trained models, a first trained model 107b and a second trained model 107c, and correlation data 107d. The storage unit 107 may be configured as one physical device, or may be distributed and arranged in multiple physical devices.

[0028] The flour milling yield estimation program 107a is a program for causing the estimation system 1 to realize the functions of the acquisition unit 101, the output unit 102, the calculation unit 103, the estimation unit 104, the derivation unit 105, and the model generation unit 106. The flour yield estimation program 107a may be written to the storage unit 107, for example, during the manufacturing stage of the estimation system 1. For example, the flour yield estimation program 107a may be distributed by another remote server device or the like and acquired by the estimation system 1 via communication. The flour yield estimation program 107a may be read by the estimation system 1 from a recording medium such as a memory card or an optical disk, and stored in the storage unit 107. The flour yield estimation program 107a may be read from a recording medium by a writing device and written to the storage unit 107 of the estimation system 1. The flour yield estimation program 107a may be provided in the form of distribution via a network, or may be provided in the form of being recorded on a recording medium.

[0029] The first trained model 107b and the second trained model 107c are trained models that have been previously subjected to machine learning or deep learning using teacher data. The trained models perform a predetermined calculation on an input value and output the calculation result. That is, data such as coefficients and thresholds of a function that specifies this calculation are stored in the storage unit 107 as the first trained model 107b and the second trained model 107c. The first trained model 107b and the second trained model 107c may be generated by the model generation unit 106 and stored in the storage unit 107, or may be generated by an external device and written in the storage unit 107. The first trained model 107b is a trained model that has been trained to output an estimated ferulic acid concentration related index distribution based on a multi-wavelength image of the endosperm region of the alkali-treated cut surface of the seed of the test grain, acquired by the image acquisition device 2. The second trained model 107c is a trained model that has been trained to output an endosperm region applicability distribution based on a multi-wavelength image of the alkali-treated cut surface of the seed of the test grain. The estimation system 1 executes the flour milling yield estimation program 107a and reads data stored as the first trained model 107b and the second trained model 107c. This enables the estimation system 1 to perform a calculation for outputting an endosperm region applicability distribution using the second trained model 107c, and a calculation for outputting an estimated ferulic acid concentration related index distribution using the first trained model 107b.

[0030] The correlation data 107d is data that indicates the correlation between the representative value and the milling yield of the test grain, derived using the representative value of a control grain with a known milling yield and the milling yield of the control grain. For example, the correlation data 107d is a calibration curve that is derived by plotting the representative values ​​and the milling yield values ​​of multiple control grains on a graph with the representative value on the vertical axis and the milling yield on the horizontal axis. The control grains may be two or more of the same kind of grains with different milling yields and representative values. From the viewpoint of the accuracy of the derived calibration curve, it is preferable to use a large number of control grains, for example, 5 or more, preferably 10 or more, more preferably 20 or more control grains.

[0031] 3 is a diagram showing a hardware configuration of the estimation system 1. The estimation system 1 includes a CPU 10a, a RAM 10b, a ROM 10c, an external memory 10d, an input unit 10e, a display unit 10f, and a communication unit 10g. The CPU 10a, the RAM 10b, the ROM 10c, the external memory 10d, the input unit 10e, the display unit 10f, and the communication unit 10g are connected to the CPU 10a via a system bus 10h.

[0032] The CPU 10a comprehensively controls each device connected to the system bus 10h. The RAM 10b functions as a main memory, a working area, etc. for the CPU 10a. The CPU 10a loads programs and the like required for executing processes from the ROM 10c or the external memory 10d to the RAM 10b, and executes the loaded programs to realize various operations. The ROM 10c and the external memory 10d store the BIOS and OS, which are control programs for the CPU 10a, as well as various programs and data required to realize the functions executed by the computer. The external memory 10d is composed of, for example, a flash memory, a hard disk, a DVD-RAM, a USB memory, or the like. The input unit 10e receives operation instructions and the like from a user, etc. The input unit 10e is composed of input devices such as, for example, an input button, a keyboard, a pointing device, a wireless remote control, a microphone, and a camera. The display unit 10f outputs data processed by the CPU 10a and data stored in the RAM 1010b, the ROM 10c, and the external memory 10d. The display unit 10f is composed of output devices such as a CRT display, an LCD, an organic EL panel, a printer, and a speaker. The communication unit 10g is an interface for connecting and communicating with an external device directly or via a network, and is configured with an interface such as a serial interface or a LAN interface.

[0033] 3. First trained model The first trained model 107b has been subjected to machine learning processing using training data so that when a multi-wavelength image is input, the model outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain with a multispectral camera. The first trained model 107b is configured as a neural network having an input layer that receives the input of the multi-wavelength image, an output layer that outputs the estimated ferulic acid concentration related index distribution of the input multi-wavelength image, and an intermediate layer in which parameters are trained using teacher data so as to input the multi-wavelength image and output the estimated ferulic acid concentration related index distribution of the input multi-wavelength image. For example, a CNN (Convolutional Neural Network) can be adopted as the first trained model 11b of this embodiment.

[0034] Hereinafter, a case where the first trained model 107b is a CNN will be described as an example. Fig. 4 is a schematic diagram for explaining the generation process of the first trained model 107b, and conceptually illustrates the process of generating the first trained model 107b by performing machine learning. The estimation system 1 generates a first trained model 107b by performing deep learning for learning the distribution of estimated ferulic acid concentration-related indices in the multi-wavelength image of the endosperm region for the CNN model. The multi-wavelength image of the endosperm region is a multi-wavelength image in which only the endosperm region of the test grain seed cut surface is cut out, and refers to an image in which regions other than the endosperm region, such as the background region, noise, and the outer shell region of the cut surface, are deleted from the multi-wavelength image of the test grain seed cut surface captured by the image acquisition device 2.

[0035] As shown in FIG. 4, the input layer of the first trained model 107b has a plurality of neurons that accept input of pixel values ​​of each pixel included in the multi-wavelength image of the endosperm region, and passes the input pixel values ​​to the intermediate layer. The intermediate layer has a plurality of neurons that extract image features, and has a configuration in which a convolution layer that convolutes the pixel values ​​of each pixel input from the input layer and a pooling layer that maps the pixel values ​​convolved in the convolution layer are alternately connected, and finally extracts image features while compressing the pixel information of the image. The intermediate layer passes the extracted features to the output layer. The output layer has one or more neurons that output an estimated ferulic acid concentration related index distribution in the multi-wavelength image of the endosperm region. The output layer outputs an estimated ferulic acid concentration related index distribution based on the image features output from the intermediate layer.

[0036] [Teacher data] In generating the first trained model 107b, a learning process is performed by acquiring a plurality of pieces of training data and using the training data. The training data includes first training data and second training data.

[0037] The first training data is data in which a first value is associated with a first multi-wavelength image created by photographing an endosperm region of an alkali-treated cut surface of a seed of a grain for use as first training data, which indicates a first milling yield, using the image acquisition device 2. That is, each pixel of the first multi-wavelength image is labeled with a first value (e.g., "-1") corresponding to the first milling yield. The second training data is data in which a second multi-wavelength image, which is created by photographing an endosperm region of an alkali-treated cut surface of a seed of a grain for use as second training data and which indicates a second milling yield, is associated with a second value. That is, each pixel of the second multi-wavelength image is labeled with a second value (e.g., "1") corresponding to the second milling yield. As the first training data grain and the second training data grain, a combination having a difference in their respective milling yields (a combination of high milling ability and low milling ability) may be adopted, or a combination having a difference in the amount of ferulic acid released when the endosperm is treated with alkali (a combination of high ferulic acid and low ferulic acid) may be adopted.

[0038] The first multi-wavelength image can be labeled with the first value manually by a user on a computer. For example, the first multi-wavelength image can be converted into a color image based on a color scale (e.g., an image on an RGB scale) and displayed on a computer screen, and the user can select the endosperm region on the screen and label it with the first value. Alternatively, the user can color the endosperm region on the screen in any color (e.g., blue), and the computer can automatically associate the color with the first value to label it. Similarly, the second multi-wavelength image can be labeled with the second value by converting the second multi-wavelength image into a color image based on a color scale and displaying it on a computer screen, and the user can select the endosperm region on the screen and label it with the second value. Alternatively, the user can color the endosperm region on the screen in any color (e.g., red), and the computer can automatically associate the color with the second value to label it.

[0039] [Distribution of estimated ferulic acid concentration-related indicators] The estimated ferulic acid concentration related index distribution is data that numerically represents the concentration of ferulic acid at each pixel of the multi-wavelength image input to the first trained model 107b. The estimated ferulic acid concentration related index distribution, for example, uses the spectral information extracted from the first multi-wavelength image and the spectral information extracted from the second multi-wavelength image as correct values, and expresses numerically whether the spectral information of each pixel of the input multi-wavelength image is closer to the first multi-wavelength image or the second multi-wavelength image. In this case, for example, a first value "1" is assigned to a pixel having spectral information closer to the first multi-wavelength image, and a second value "-1" is assigned to a pixel having spectral information closer to the second multi-wavelength image. Alternatively, the estimated ferulic acid concentration related index distribution may express the spectral information of each pixel as a continuous probability value. In this case, for example, a first value "1" is assigned to a pixel having spectral information closest to the first multi-wavelength image, a second value "-1" is assigned to a pixel having spectral information closest to the second multi-wavelength image, a value "2" is assigned to a pixel having spectral information most deviating from the first multi-wavelength image side, and a value "-2" is assigned to a pixel having spectral information most deviating from the second multi-wavelength image side, and each pixel is assigned a continuous value from "-2" to "2" according to its spectral information. The value assigned to each pixel corresponds to the milling yield, but also indirectly corresponds to the concentration of ferulic acid because the milling yield correlates with the concentration of ferulic acid in the endosperm region. In other words, the estimated ferulic acid concentration related index distribution is not the value of the amount of ferulic acid in each pixel itself, but is an index based on the amount of ferulic acid in each pixel.

[0040] 5. Estimation of flour milling yield Hereinafter, the estimation of the flour milling yield of the test grain executed by the estimation system 1 will be described with reference to the flowchart of FIG.

[0041] The acquisition unit 101 of the estimation system 1 communicates with the image acquisition device 2 via the communication unit 10g, and acquires a multi-wavelength image of the endosperm region, which is created by photographing the endosperm region of the cut surface of the seed of the test grain that has been subjected to the alkali treatment, using a multispectral camera, acquired by the image acquisition device 2 (step S11). The acquisition unit 101 stores the acquired multi-wavelength image in the storage unit 107.

[0042] The output unit 102 inputs the multi-wavelength image of the endosperm region acquired in step S11 to the first trained model 107b stored in the memory unit 107 (step S12). The output unit 102 causes the first trained model 107b to output an estimated ferulic acid concentration related index distribution (step S13). The output unit 102 causes the memory unit 107 to store the estimated ferulic acid concentration related index distribution.

[0043] The calculation unit 103 calculates a representative value based on the estimated ferulic acid concentration related index distribution output in step S13 (step S14). The representative value is, for example, an average value obtained by dividing the sum of the values ​​labeled to each pixel on the distribution of the estimated ferulic acid concentration-related index by the total number of pixels. Alternatively, the representative value may be the mode of the values ​​labeled to each pixel on the distribution of the estimated ferulic acid concentration-related index, or may be the ratio of the number of pixels of each value after binarizing the value corresponding to the ferulic acid concentration assigned to each pixel using a threshold value. The calculation unit 103 stores the representative value in the storage unit 107.

[0044] The derivation unit 105 creates the correlation data 107d (step S15). The process of step S15 will be described with reference to FIG. The derivation unit 105 acquires multi-wavelength images of the endosperm region of the alkaline-treated cut surface of seeds of multiple control grains with known milling yields (step S151), inputs the multi-wavelength images of the endosperm region to the first trained model 107b (step S152), causes the first trained model 107b to output an estimated ferulic acid concentration related index distribution (step S153), and calculates a representative value (step S154).

[0045] The derivation unit 105 plots the representative value of the control grain and the milling yield calculated in step S154 on a graph with the vertical axis representing the representative value and the horizontal axis representing the milling yield to create a calibration curve (step S155). The derivation unit 105 stores the calibration curve in the memory unit 107 as correlation data 107d.

[0046] When the generation of the correlation data 107d is completed, the estimation unit 104 refers to the correlation data 107d and the representative value of the test grain calculated in step S14 to estimate the milling yield of the test grain (step S16). The estimation unit 104 may store the obtained milling yield in the memory unit 107, or display it on the display unit 10f included in the estimation system 1. With the above, the estimation system 1 ends the processing.

[0047] According to the present invention, it is possible to predict the milling yield without milling the test grain. In addition, it is not necessary to actually measure the content of ferulic acid in the test grain. That is, according to the method of this embodiment, it is possible to easily estimate the milling yield from a small amount (e.g., 10 grains or less) of test grain seeds.

[0048] In the second embodiment, the estimation of the flour milling yield is performed by the estimation system 1, but this is not limited to the above, and the estimation may be performed by the user. In this case, for example, the correlation data 107d is displayed on the display unit 10f of the estimation system 1, and the user can refer to the correlation data 107d to determine the flour milling yield based on the representative value of the test grain.

[0049] <Second aspect> The second aspect of the second embodiment will now be described. In the second embodiment, the estimation system 1 acquires a multi-wavelength image of the endosperm region by identifying the endosperm region from the multi-wavelength image using the second trained model 107c. The other configurations are the same as those in the first embodiment, and therefore the description will be omitted.

[0050] 1. Second trained model The second trained model 107c has been subjected to machine learning processing using training data so that when a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of a test grain with a multispectral camera is input, the model outputs a distribution of endosperm region relevance of the multi-wavelength image. The second trained model 107c is configured as a neural network having an input layer that accepts the input of the multi-wavelength image, an output layer that outputs the endosperm region applicability distribution of the input multi-wavelength image, and an intermediate layer in which parameters are trained using teacher data so as to input the multi-wavelength image and output the endosperm region applicability distribution of the input multi-wavelength image. For example, CNN can be adopted as the second trained model 107c of this embodiment.

[0051] [Teacher data] In generating the second trained model 107c, a learning process is performed using the third training data. The third training data is data in which the pixels in the endosperm region are labeled with a third value (e.g., "1") and the pixels outside the endosperm region are labeled with a fourth value (e.g., "-1") in a third multi-wavelength image created by photographing an alkali-treated cut surface of a seed of a grain for the third training data by the image acquisition device 2. Note that the first multi-wavelength image and / or the second multi-wavelength image used in the training data of the first trained model 11b can also be used as the third multi-wavelength image.

[0052] The third multi-wavelength image can be labeled with the third value and the fourth value manually by a user on a computer. For example, the third multi-wavelength image can be converted into a color image based on a color scale and displayed on a computer screen, and the user can select the endosperm region on the screen and label it with the third value, and label the region other than the endosperm region with the fourth value. Alternatively, the user can color the endosperm region on the screen in any color (e.g., green), color the region other than the endosperm region in another color (e.g., yellow), and the computer can automatically associate each color with the third value or the fourth value for labeling.

[0053] [Endosperm region applicable distribution] The endosperm region applicability distribution is data that numerically represents whether or not each pixel of the multi-wavelength image input to the second trained model 107c is an endosperm region. The endosperm region applicability distribution, for example, takes the spectral information extracted from the endosperm region and the region other than the endosperm region of the third multi-wavelength image as the correct answer value, and expresses numerically which region of the third multi-wavelength image the spectral information of each pixel of the input multi-wavelength image is close to. In this case, for example, a third value "1" is assigned to pixels having spectral information close to the endosperm region of the third multi-wavelength image, and a fourth value "-1" is assigned to pixels having spectral information close to the region other than the endosperm region of the third multi-wavelength image. Alternatively, the endosperm region applicability distribution may express the spectral information of each pixel as a continuous probability value. In this case, for example, a third value "1" is assigned to a pixel having spectral information closest to the endosperm region of the third multi-wavelength image, a fourth value "-1" is assigned to a pixel having spectral information closest to a region other than the endosperm region of the third multi-wavelength image, a value "2" is assigned to a pixel having spectral information most deviating from the endosperm region side, and a value "-2" is assigned to a pixel having spectral information most deviating from the region other than the endosperm region side, and each pixel is assigned a continuous value from "-2" to "2" according to its spectral information.

[0054] 2. Estimation of flour milling yield Hereinafter, the estimation of the flour milling yield of the test grain executed by the estimation system 1 will be described with reference to the flowchart of FIG.

[0055] The acquisition unit 101 of the estimation system 1 communicates with the image acquisition device 2 via the communication unit 10f, and acquires a multi-wavelength image of the cut surface of the seed of the test grain that has been subjected to the alkali treatment, which is acquired by the image acquisition device 2 (step S21). The multi-wavelength image includes an area other than the endosperm area. The acquisition unit 101 stores the acquired multi-wavelength image in the memory unit 107. The acquisition unit 101 inputs the multi-wavelength image acquired in step S21 to the second trained model 107c stored in the storage unit 107 (step S22). The acquisition unit 101 causes the second trained model 107c to output the endosperm region applicability distribution (step S23), and extracts a multi-wavelength image of the endosperm region of the cut surface of the seed of the test grain based on the endosperm region applicability distribution (step S24). In step S24, the values ​​assigned to each pixel in the endosperm region applicability distribution are binarized using a threshold value to identify the endosperm region of the multi-wavelength image. For example, when the endosperm region applicability distribution is such that each pixel is assigned a continuous value from "2" to "-2", each value is binarized to "1" or "-1" using a threshold value "0" to determine whether each pixel is an endosperm region (e.g., a value of "1" is assigned) or a region other than the endosperm region (e.g., a value of "-1" is assigned). The acquisition unit 101 cuts out the region identified as the endosperm region from the multi-wavelength image, sets it as a multi-wavelength image of the endosperm region, and stores it in the storage unit 107.

[0056] The output unit 102 inputs the multi-wavelength image of the endosperm region acquired in step S24 to the first trained model 107b stored in the storage unit 107 (step S25). The processes of steps S26 and S27 are similar to the processes of steps S13 and S14 in FIG.

[0057] The derivation unit 105 creates the correlation data 107d (step S28). The process of step S28 will be described with reference to FIG. The derivation unit 105 communicates with the image acquisition device 2 via the communication unit 10f, acquires multi-wavelength images of the cut surfaces of seeds of multiple control grains that have been subjected to alkali treatment acquired by the image acquisition device 2 (step S281), inputs the multi-wavelength images to the second trained model 107c (step S282), causes the second trained model 107c to output an endosperm region applicability distribution (step S283), and extracts multi-wavelength images of the endosperm region of the cut surfaces of the seeds of the control grain based on the endosperm region applicability distribution (step S284). The output unit 102 inputs the multi-wavelength image of the endosperm region acquired in step S284 to the first trained model 107b stored in the storage unit 107 (step S285). The processes of steps S286 to S288 are similar to the processes of steps S153 to S155 in FIG. As described above, by using the second trained model 107c, the endosperm region can be automatically extracted from the multi-wavelength image and used to estimate the milling yield.

[0058] <Third aspect> The third aspect of the second embodiment will now be described. In the third aspect, the estimation system 1 generates a first trained model 107b and a second trained model 107c. The other configurations are similar to those in the first aspect, and therefore will not be described.

[0059] 1. Generation of the first trained model The estimation system 1 generates a first trained model 107b by the model generation unit 106 using the first training data and the second training data. The model generation unit 106 inputs the first teacher data and the second teacher data to the input layer of the CNN, and obtains an estimated ferulic acid concentration related index distribution output from the output layer after arithmetic processing in the intermediate layer. In the intermediate layer of the CNN, spectral information of each pixel is extracted from the images of multiple wavelength bands included in the first multi-wavelength image and the second multi-wavelength image, and a numerical value representing the ferulic acid concentration at each pixel is calculated.

[0060] The model generation unit 106 compares the estimated ferulic acid concentration related index distribution output from the output layer with information labeled for the first multi-wavelength image or the second multi-wavelength image, i.e., the correct answer value (for example, the first value "-1" or the second value "1"), and optimizes the parameters used in the calculation process in the intermediate layer so that the output value from the output layer approaches the correct answer value. The parameters are, for example, weights (coupling coefficients) between neurons. There are no particular limitations on the method of optimizing the parameters, and for example, the model generation unit 106 can optimize various parameters using the steepest descent method or the like. For example, the parameters may be modified so that the spectral information possessed by each of the first multi-wavelength image and the second multi-wavelength image is weighted to images of wavelengths that are significantly distinguishable. The model generation unit 106 repeatedly performs the above process for each of the first teacher data and the second teacher data, thereby obtaining a first trained model 107b for which the learning process has been completed.

[0061] 9 is a flowchart showing a method for generating the first trained model 107b. The model generation unit 106 acquires, for example, first teacher data and second teacher data prepared in advance and stored in the storage unit 107 (step S31). The model generation unit 106 generates the first trained model 107b having, for example, a CNN configuration, using the acquired first teacher data and second teacher data (step S32). After the end of learning, the model generation unit 106 stores the generated first trained model 107b in the storage unit 107 (step S33). With the above, the generation process of the first trained model 107b is completed.

[0062] 2. Generate a second trained model The estimation system 1 generates, by the model generation unit 106, a second trained model 107c using the third training data. The model generation unit 106 inputs the third training data to the input layer of the CNN, and obtains an endosperm region applicability distribution output from the output layer after arithmetic processing in the intermediate layer. In the intermediate layer of the CNN, spectral information of each pixel is extracted from the images of multiple wavelength bands included in the third multi-wavelength image, and a numerical value indicating whether each pixel is an endosperm region or not is calculated.

[0063] The model generation unit 106 compares the endosperm region applicability distribution output from the output layer with information labeled for the third multi-wavelength image, i.e., the correct answer value (e.g., the third value "1" or the fourth value "-1"), and optimizes parameters used in the calculation process in the intermediate layer so that the output value from the output layer approaches the correct answer value. The parameters are, for example, weights (coupling coefficients) between neurons. The method for optimizing the parameters is not particularly limited, and for example, the model generation unit 106 can optimize various parameters using the steepest descent method or the like. For example, the parameters may be modified so that the spectral information possessed by the endosperm region and the region other than the endosperm region of the third multi-wavelength image is weighted to an image of a wavelength that can be significantly identified.

[0064] 10 is a flowchart showing a method for generating the second trained model 107c. The model generation unit 106 acquires, for example, third training data prepared in advance and stored in the storage unit 12 (step S41). The model generation unit 106 uses the acquired third training data to generate the second trained model 107c having, for example, a CNN configuration (step S42). After the end of learning, the model generation unit 106 stores the generated second trained model 107c in the storage unit 107 (step S43). This completes the process for generating the second trained model 107c.

[0065] Using the first trained model 107b and the second trained model 107c generated as described above, it is possible to perform the processing described in the first aspect or the second aspect. EXAMPLES

[0066] The present invention will be specifically described below with reference to examples, but the scope of the present invention is not limited to these examples.

[0067] <Test Example 1: Correlation between milling yield and free ferulic acid> A correlation was confirmed between wheat flour yield and free ferulic acid. 1.5 kg of seeds were milled and milling yield was calculated. The moisture content was adjusted (water was added) to 14.5% 24 hours before milling, and milling was performed in accordance with the AACC26 method using a test mill (Bhullar test mill; manufactured by Bhullar Co., Ltd.). After milling, wheat flour was mixed in the order of B1+M1, B2+M2, B3, and M3 to prepare "60% flour" that reached 60% of the total amount. Milling yield was calculated using the following formula. Milling yield (%) = flour / (flour + bran) x 100

[0068] Ferulic acid contained in the endosperm of the seeds was liberated and quantified. Specifically, 5 ml of 1M NaOH was added to 100 mg of the above 60% powder and mixed by inversion for 12 hours. Then, 1.25 ml of 3M HCl and 4.75 ml of 0.5M glycine hydrochloride buffer (pH 2.0) were added to 4 ml of the supernatant and mixed. Further, 10 ml of ethyl acetate was added and shaken. 15 ml of the solution was concentrated to dryness with nitrogen and dissolved in 50% methanol. Ferulic acid was quantified by HPLC with modifications to Harukaze et al. (1999). In Figure 1, the measured values ​​are plotted with milling yield on the vertical axis and free ferulic acid concentration on the horizontal axis, and a calibration curve was drawn in Excel. The correlation coefficient (r) in Figure 1 was calculated as Pearson's product moment correlation coefficient using R version 4.2.0. As shown in Figure 1, it was confirmed that there is a negative correlation between milling yield and free ferulic acid concentration. If the milling yield of a test grain is unknown, the milling yield can be estimated by measuring the free ferulic acid concentration and applying it to the calibration curve.

[0069] <Test Example 2: Estimation of flour milling yield using multi-wavelength images> 1.Equipment used In this test example, acquisition of multi-wavelength images and estimation of flour milling yield were performed using VideometerLab 4 (manufactured by Videometer Inc.). In this test example, VideometerLab 4 was made to function as the estimation system 1 and the image acquisition device 2.

[0070] 2. Acquiring multi-wavelength images for creating training data As the first training data grain seeds, 60 seeds of the known high milling quality variety "China 175" (sown in 2018, harvested in 2019, milling yield 71.5%) were used. As the second training data grain seeds, 60 seeds of the known low milling quality variety "Shirogane Wheat" (sown in 2018, harvested in 2019, milling yield 66.0%) were used. Each seed was cut approximately in half to prepare half-cut seeds. The half-cut seeds were attached to a petri dish with double-sided tape with the cross-cut surface facing up. 4 μl of 0.1 M ammonia water was dropped onto the cut surface and left to stand for 30 minutes to release ferulic acid.

[0071] The ammonia-treated first training data grain and second training data grain were irradiated with 19 wavelengths (365nm, 405nm, 430nm, 450nm, 470nm, 490nm, 515nm, 540nm, 570nm, 590nm, 630nm, 645nm, 660nm, 690nm, 780nm, 850nm, 880nm, 940nm, and 970nm) in the wavelength band of 365nm to 970nm using VideometerLab 4 to obtain a multi-wavelength image. Figure 11 shows an image obtained by converting the multi-wavelength image into an RGB scale.

[0072] 3. Creating a second trained model The third training data was created using the multi-wavelength image of the grain for the first training data and the multi-wavelength image of the grain for the second training data as a third multi-wavelength image. FIG. 12A shows third training data created by labeling the third multi-wavelength image with a third value and a fourth value. For an image obtained by converting the multi-wavelength image shown in the left diagram of FIG. 12A into an RGB scale, an experimenter manually colored the endosperm region of the third multi-wavelength image green and the region other than the endosperm region yellow on a computer screen, as shown in the right diagram of FIG. 12A. At this time, the endosperm region of the third multi-wavelength image was labeled with "1" corresponding to green, and the region other than the endosperm region was labeled with "-1" corresponding to yellow. A model "product vs. background" that distinguishes between the endosperm region and the region other than the endosperm region was generated using the software "Transformation builder" of VideometerLab 4 for the third training data created in this way. In this test example, the model "product vs. background" functions as a second trained model, and henceforth, the model will be referred to as the second trained model.

[0073] FIG. 12B shows the result of applying the second trained model to the third multi-wavelength image. The image shown in FIG. 12B shows the endosperm region applicability distribution converted to an RGB scale. In the endosperm region applicability distribution, "1" is assigned to the pixel having the spectral information closest to the endosperm region (pixels colored green in the right diagram of FIG. 12A), "-1" is assigned to the pixel having the spectral information closest to the region other than the endosperm region (pixels colored yellow in the right diagram of FIG. 12A), "2" is assigned to the pixel having the spectral information most deviating from the positive side, and "-2" is assigned to the pixel having the spectral information most deviating from the negative side, and each pixel is assigned a continuous value from "-2" to "2" according to its spectral information. In FIG. 12B, the assigned values ​​are converted to colors ranging from red (corresponding to "2") to blue (corresponding to "-2") and displayed.

[0074] 4. Extraction of Endosperm Regions From the image shown in Fig. 12B, the endosperm region was extracted by binarizing the values ​​labeled in the pixels using a threshold value of "0". A model "Segmentation" for extracting the endosperm region was generated using the software "Segmentation builder" of VideometerLab 4. The specific "Segmentation builder" settings are as follows: Transformation: Select "product vs. background" Type: Simple Threshold Morphology Threshold: 0 Morphology Filter: Euclidean Open Disk Radius: 7 Fig. 12C shows an image in which the endosperm region has been extracted after binarization. In Fig. 12C, the endosperm region is shown in white and the region other than the endosperm region is shown in black, and noise such as dust reflected in the multi-wavelength image has also been removed.

[0075] 5. Creating the first trained model The multi-wavelength image of the grain for the first training data and the multi-wavelength image of the grain for the second training data were input to the second trained model to output the endosperm region appropriateness distribution, and the endosperm region was extracted by binarizing the endosperm region appropriateness distribution to create the first multi-wavelength image and the second multi-wavelength image. Figure 13A shows images obtained by converting the multi-wavelength image of the grain for the first training data and the multi-wavelength image of the grain for the second training data before extracting the endosperm region into RGB scale. Figure 13B shows images obtained by converting the first multi-wavelength image and the second multi-wavelength image created by extracting the endosperm region into RGB scale. Next, the first and second multi-wavelength images were labeled with a first value and a second value, respectively. As shown in FIG. 14A, the endosperm region of the first multi-wavelength image was colored light blue, and the endosperm region of the second multi-wavelength image was colored dark red. At this time, the endosperm region of the first multi-wavelength image was labeled with "-1" corresponding to light blue, and the endosperm region of the second multi-wavelength image was labeled with "1" corresponding to dark red. Next, a model "seed FA analysis" that outputs an estimated ferulic acid concentration related index distribution was generated using the software "Transformation builder" of VideometerLab 4. In this test example, the model "seed FA analysis" functions as a first trained model, and therefore the model is described as a first trained model.

[0076] FIG. 14B shows the result of applying the first trained model to the first multi-wavelength image and the second trained model. The image shown in FIG. 14B shows the estimated ferulic acid concentration related index distribution converted to an RGB scale. In the estimated ferulic acid concentration related index distribution, "-1" is assigned to the pixel having the closest spectral information to the pixel colored light blue in the first multi-wavelength image, "1" is assigned to the pixel having the closest spectral information to the pixel colored dark red in the second multi-wavelength image, "2" is assigned to the pixel having the most positive spectral information, and "-2" is assigned to the pixel having the most negative spectral information, and each pixel is assigned a continuous value from "-2" to "2" according to the spectral information. In FIG. 14B, each pixel is colored in a color ranging from dark red (corresponding to "2") to dark blue (corresponding to "-2") according to the assigned value.

[0077] 6. Calculation of representative value Representative values ​​for the control grains were calculated using the generated first and second trained models. As control grains, 19 varieties and lines sown in 2018 and 2019 were used, which were Japanese noodle lines (F6-7 generation) developed at the West Japan Agricultural Research Center of the National Agriculture and Food Research Organization. Ten seeds of each variety and line were used, and multi-wavelength images of the cut surfaces of the seeds were obtained. The multi-wavelength images were obtained in the same manner as in "2. Acquiring multi-wavelength images for creating training data" above.

[0078] From the multi-wavelength images, representative values ​​were calculated for each variety and line used as the control grain. Using the VideometerLab 4 software "Session Manager", we designed a model that extracts the endosperm region from the multi-wavelength image of the control grain using the second trained model, outputs the estimated ferulic acid concentration related index distribution from the multi-wavelength image of the endosperm region using the first trained model, and calculates a representative value from the estimated ferulic acid concentration related index distribution. In this test example, the representative value is the average value obtained by dividing the sum of the values ​​(values ​​from "-2" to "2") labeled to each pixel of the estimated ferulic acid concentration related index distribution by the total number of pixels. The specific "Session Manager" settings are as follows: Plugin Algorism: General Statistic Name: Statistic Statistics: Select "Segmentation (created in "3. Creating the second trained model" above)" and "seed FA analysis". The representative values ​​obtained are shown in Table 1 below.

[0079] 7. Creating a calibration curve In order to create a calibration curve using the representative value and milling yield of the control grain, each variety and line of the control grain was milled and the milling yield was measured. 1.5 kg of seeds of each variety and line were used for milling. The moisture content was adjusted (water was added) to 14.5% 24 hours before milling, and milling was performed according to the AACC26 method using a test flour mill (Bhullar test mill; manufactured by Bhuller Co., Ltd.). After milling, wheat flours were mixed in the order of B1+M1, B2+M2, B3, and M3 to prepare "60% flour" that reached 60% of the total amount. The milling yield was calculated using the following formula. Milling yield (%) = flour / (flour + bran) x 100 The milling yields obtained are shown in Table 1.

[0080] [Table 1]

[0081] As shown in Figure 15, a scatter plot was created by plotting the representative value of each control grain on the vertical axis and the milling yield value on the horizontal axis, and an approximation curve for the scatter plot was drawn in Excel. The approximation curve functions as a calibration curve. The correlation coefficient (r) in Figure 15 was calculated as Pearson's product moment correlation coefficient using R version 4.2.0. As is clear from Figure 15, there is a negative correlation between the milling yield and the representative value. By referring to the calibration curve shown in FIG. 15, the milling yield of the test grain can be predicted based on the representative value of the test grain whose milling yield is unknown.

[0082] 8. Correlation between ferulic acid content and representative value Ferulic acid in the seeds of the control grains was quantified. The above 60% flour was used for the quantification of ferulic acid. The quantification of ferulic acid was performed by HPLC with modifications to Harukaze et al. (1999). The ferulic acid (FA) concentrations obtained are shown in Table 1. In Fig. 16, a scatter diagram is created by plotting the values ​​of the control grains with the vertical axis representing the representative value and the horizontal axis representing the ferulic acid (FA) concentration, and an approximation curve is drawn for the scatter diagram. The approximation curve functions as a calibration curve. As is clear from Fig. 16, there is a positive correlation between the ferulic acid concentration and the representative value.

[0083] 9. Arabinoxylan content and correlation with representative values Arabinoxylan contained in the seeds of the control grains was quantified. The above 60% flour was used for arabinoxylan quantification. Arabinoxylan quantification was performed by GC according to Shewry and Ward edited (2016). The obtained arabinoxylan (AX) concentrations are shown in Table 1. Figure 17 is a scatter diagram created by plotting the values ​​of the control grains with the vertical axis representing the representative value and the horizontal axis representing the arabinoxylan (AX) concentration, and then drawing an approximation curve for the scatter diagram. The approximation curve functions as a calibration curve. As is clear from Figure 17, there is a positive correlation between the arabinoxylan concentration and the representative value. The results of Test Example 2 showed that the distribution of the estimated ferulic acid concentration-related index indirectly represents the amount of ferulic acid, and that the milling yield can be predicted using the representative value calculated from the distribution of the estimated ferulic acid concentration-related index. [Explanation of symbols]

[0084] 1. Estimation System 101 Acquisition Department 102 Output section 103 Calculation section 104 Estimation part 105 Derivation part 106 Model Generation Unit 107 Storage section 107a Milling yield estimation program 107b First trained model 107c Second trained model 107d Correlation Data 10a CPU 10b RAM 10c ROM 10d External Memory 10e Input section 10f display section 10g Communication section 10h system bus 2. Image acquisition device N Network

Claims

1. 1. A method for estimating milling yield of a test grain, comprising: (i) cutting the seeds of the test grain and treating the cut surface with alkali; (ii) measuring a value corresponding to the amount of ferulic acid in the endosperm region of the alkali-treated cut surface of the seed of the test grain; (iii) estimating the milling yield of the test grain from the value measured in step (ii) using a calibration curve calculated using a plurality of milling yield data of the control grain and values ​​corresponding to the amount of ferulic acid measured in the endosperm region of the alkali-treated cut surface of the seed of the control grain; A method comprising:

2. 2. The method of claim 1, wherein the value corresponding to the amount of ferulic acid is based on the absorbance attributable to ferulic acid.

3. The method of claim 1 , wherein the alkali treating step comprises dripping 0.1 M aqueous ammonia onto the cross section.

4. (1) The test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; The method of claim 1.

5. 1. A computer-implemented method for estimating milling yield of a test grain, comprising: Obtaining a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the test grain seed using a multispectral camera; inputting the multi-wavelength image into a first trained model and outputting an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; Calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; Estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined; Including, The first trained model is a model that has been subjected to machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the first trained model outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

6. 1. A method for estimating milling yield of a test grain, comprising: (i) Obtaining a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the test grain seed using a multispectral camera; inputting the multi-wavelength image into a first trained model and outputting an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; Calculating a representative value corresponding to the ferulic acid concentration in the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; (ii) estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined; Including, The first trained model is a model that has been subjected to machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the first trained model outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

7. The teacher data is First training data in which a first multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for first training data, which shows a first milling yield, with a multispectral camera, and a first value is associated with the first training data; the second training data including a second multi-wavelength image produced by photographing an endosperm region of an alkali-treated cut surface of a grain seed for use as second training data, which indicates a second milling yield, with a multispectral camera, and a second value associated therewith; the distribution of indexes related to the estimated ferulic acid concentration of the multi-wavelength image is data expressed as a value corresponding to the ferulic acid concentration estimated for each pixel of the multi-wavelength image, The method according to claim 5 or 6, wherein calculating the representative value is an average value obtained by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels.

8. Obtaining a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera Obtaining a multi-wavelength image by photographing the alkali-treated cut surface of the test grain seed with a multispectral camera; The multi-wavelength image is input to a second trained model, and an endosperm region applicability distribution of the multi-wavelength image is output; Extracting a multi-wavelength image of the endosperm region of the cross section of the seed of the test grain based on the endosperm region relevance distribution; The method described in claim 5 or 6, wherein the second trained model has undergone machine learning processing using training data so that when a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the second trained model outputs an endosperm region relevance distribution of the multi-wavelength image.

9. and further comprising deriving a correlation between the representative value and the milling yield of the test grain; Deriving the correlation may include: Acquiring multi-wavelength images of the endosperm regions of the seeds of a plurality of control grains with known milling yields by photographing the endosperm regions of the alkali-treated cut surfaces with a multispectral camera; The multi-wavelength image is input to the first trained model, and an estimated ferulic acid concentration-related index distribution of the multi-wavelength image is output; Calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; and creating a calibration curve using a plurality of the calculated representative values ​​and the milling yields of the control grains; 7. The method according to claim 5 or 6.

10. The method according to claim 5 or 6, wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

11. (1) The test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; 7. The method according to claim 5 or 6.

12. A system for estimating an index distribution related to an estimated ferulic acid concentration in an endosperm region of a cut surface of a seed of a test grain that has been treated with alkali, comprising: An acquisition unit that acquires a multi-wavelength image of the endosperm region by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera; an output unit that inputs the multi-wavelength image to a first trained model and outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; a calculation unit that calculates a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The first trained model is a system that has undergone machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the system outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

13. 1. A system for estimating milling yield of a test grain, comprising: An acquisition unit that acquires a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera; an output unit that inputs the multi-wavelength image to a first trained model and outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; a calculation unit that calculates a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The first trained model is a system that has undergone machine learning processing using training data so that when a multi-wavelength image created by photographing the endosperm region of an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the system outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image.

14. The teacher data is First training data in which a first multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for first training data indicating a first milling yield using a multispectral camera, and a first value is associated with the first training data; and second training data in which a second multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for second training data indicating a second milling yield, and a second value is associated with the second multi-wavelength image, the second multi-wavelength image being created by photographing an endosperm region of an alkali-treated cut surface of a grain seed using a multispectral camera, and the second training data indicates a second milling yield. the distribution of indices related to the estimated ferulic acid concentration of the multi-wavelength image is data in which a value corresponding to an estimated ferulic acid concentration is expressed for each pixel of the multi-wavelength image, The system according to claim 12 or 13, wherein the calculation unit calculates an average value by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels as the representative value.

15. The acquisition unit A multi-wavelength image is obtained by photographing the alkali-treated cut surface of the seed of the test grain using a multispectral camera; The multi-wavelength image is input to a second trained model, and an endosperm region applicability distribution of the multi-wavelength image is output; Based on the endosperm region relevance distribution, a multi-wavelength image of the endosperm region of the cross section of the seed of the test grain is extracted, wherein: The system described in claim 12 or 13, wherein the second trained model has undergone machine learning processing using training data so that when a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera is input, the second trained model outputs a distribution of endosperm region relevance of the multi-wavelength image.

16. The system described in claim 12 or 13 further comprises an estimation unit that estimates the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined.

17. Further provided is a derivation unit that derives a correlation between the representative value and the milling yield of the test grain, the acquisition unit acquires multi-wavelength images of the endosperm regions prepared by photographing, with a multispectral camera, endosperm regions of alkali-treated cut surfaces of seeds of a plurality of control grains having known milling yields; the output unit inputs the multi-wavelength image to the first trained model and outputs an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; The calculation unit calculates a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution, The system according to claim 12 or 13, wherein the deriving unit creates a calibration curve using a plurality of the calculated representative values ​​and the milling yields of the control grains.

18. The system according to claim 12 or 13, wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

19. (1) The test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The test grain is wheat; 14. A system according to claim 12 or 13.

20. On the computer, A process of acquiring a multi-wavelength image by photographing the endosperm region of the alkali-treated cut surface of the test grain seed using a multispectral camera; A process of inputting the multi-wavelength image to a first trained model and outputting an estimated ferulic acid concentration-related index distribution of the multi-wavelength image; A process of calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; A computer program for causing the execution of The first trained model is a model that has been subjected to machine learning processing using training data so as to output an estimated ferulic acid concentration-related index distribution of a multi-wavelength image created by photographing an endosperm region of an alkali-treated cut surface of a seed of the test grain with a multispectral camera when the multi-wavelength image is input. Computer program.

21. The teacher data is First training data in which a first multi-wavelength image of an endosperm region of an alkali-treated cut surface of a grain seed for first training data indicating a first milling yield is associated with a first value; and second training data in which a second multi-wavelength image of an endosperm region of an alkali-treated cut surface of a grain seed for second training data indicating a second milling yield is taken by a multispectral camera and a second value is associated with the second training data, the distribution of indices related to the estimated ferulic acid concentration of the multi-wavelength image is data in which a value corresponding to an estimated ferulic acid concentration is expressed for each pixel of the multi-wavelength image, The computer program according to claim 20, wherein the process of calculating the representative value is to calculate an average value obtained by dividing the sum of values ​​corresponding to the ferulic acid concentration labeled for each pixel output from the first trained model by the total number of pixels.

22. The acquiring process includes: A process of acquiring a multi-wavelength image by photographing the alkali-treated cut surface of the test grain seed using a multispectral camera; A process of inputting the multi-wavelength image to a second trained model and outputting an endosperm region applicability distribution of the multi-wavelength image; and extracting a multi-wavelength image of the endosperm region of the cross section of the seed of the test grain based on the endosperm region relevance distribution; The computer program of claim 20, wherein the second trained model has undergone machine learning processing using training data so as to output a distribution of endosperm region relevance of a multi-wavelength image created by photographing an alkali-treated cut surface of the seed of the test grain with a multispectral camera when the multi-wavelength image is input.

23. The computer program of claim 20, further causing the computer to execute a process of estimating the milling yield of the test grain based on the calculated representative value and a correlation between the representative value and the milling yield of the test grain that has been previously determined.

24. causing the computer to further execute a process of deriving a correlation between the representative value and the milling yield of the test grain; The deriving process includes: A process of acquiring multi-wavelength images of the endosperm region by photographing the endosperm region of an alkali-treated cut surface of seeds of a plurality of control grains having known milling yields using a multispectral camera; A process of inputting the multi-wavelength image to the first trained model and outputting an estimated ferulic acid concentration related index distribution of the multi-wavelength image; A process of calculating a representative value corresponding to the ferulic acid concentration of the entire endosperm region based on the output estimated ferulic acid concentration related index distribution; A process of creating a calibration curve using a plurality of the calculated representative values ​​and the milling yield of the control grain; 21. A computer program according to claim 20.

25. An input layer to which a multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera is input; an output layer that outputs the estimated ferulic acid concentration related index distribution of the input multi-wavelength image; an intermediate layer in which parameters are learned using training data so as to receive the multi-wavelength image as an input and output the estimated ferulic acid concentration related index distribution of the input multi-wavelength image; A computer program that causes a computer to perform a calculation based on parameters of the intermediate layer on the multi-wavelength image input to the input layer of a trained model comprising the above, and to output an estimated ferulic acid concentration-related index distribution of the multi-wavelength image from the output layer.

26. The computer program according to claim 20 or 25, wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

27. (1) The test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The computer program according to claim 20 or 25, wherein the test grain is wheat.

28. A method for generating a trained model for estimating an estimated ferulic acid concentration-related index distribution in an endosperm region of an alkali-treated cut surface of a test grain seed, comprising: Acquiring training data in which a multi-wavelength image is created by photographing an endosperm region of an alkali-treated cut surface of a grain seed for training data using a multispectral camera and the estimated ferulic acid concentration-related index distribution of the multi-wavelength image is associated with each other; When a multi-wavelength image created by photographing the endosperm region of the alkali-treated cut surface of the seed of the test grain using a multispectral camera is input, a trained model is generated based on the training data, which outputs an estimated ferulic acid concentration-related index distribution of the input multi-wavelength image; A method comprising:

29. The method according to claim 28, wherein the wavelengths used for imaging by the multispectral camera include at least one wavelength between 365 nm and 940 nm.

30. (1) The test grain contains arabinoxylan in the endosperm of the seed of the test grain, and / or (2) The method according to claim 28, wherein the test grain is wheat.

Images