Seed trait estimation device, seed trait estimation model generation device, seed trait estimation method, seed trait estimation model generation method, and program

Abstract

To provide a seed trait estimation device and a seed trait estimation method which can be used to estimate seed traits.SOLUTION: A seed trait estimation device 3 disclosed herein includes a first acquisition unit 31 that acquires a multiwavelength image in which one or more seeds are captured, and an estimation unit 32 that uses a first seed trait estimation model to estimate the traits of one or more pixels constituting a seed in the multiwavelength image, thereby estimating the seed traits.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a seed trait estimation device, a seed trait estimation model generation device, a seed trait estimation method, a seed trait estimation model generation method, and a program.

Background Art

[0002] Seed germination, genetic purity of seeds, and seed contamination can be confirmed by germination tests, biochemical tests, cultivation tests, or pathological tests. In the germination test, the seeds to be tested are made to absorb moisture, and the traits of the seeds such as the presence or absence of germination are tested by checking the roots, hypocotyls, or leaves. In the biochemical test, DNA is extracted from the seeds to be tested, and the genetic traits are tested. Also, in the cultivation test, it is tested by observing the seedlings and green fruits after cultivation. In the pathological test, the seeds to be tested are cultured, or DNA or RNA is extracted from the seeds to be tested, and pathogens are detected by performing PCR tests or the like. Also, in the pathological test, it is tested by the presence or absence of disease after cultivation. However, in each test method, since the seeds to be tested are sown or destroyed, the seeds to be tested cannot be sold as seeds. Therefore, sampling inspection is basically used for the inspection regarding the traits of the seeds.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In industrial products and the like, attempts have been made to improve the efficiency of quality inspection by using a learned model for quality inspection. Specifically, an image including the industrial product to be inspected and the quality of the inspection target are combined as teacher data, and a learned model is generated by machine learning. Then, using the obtained learned model and the image including the industrial product to be inspected, the quality of the inspection target is estimated and attempted to be utilized for quality inspection.

[0004] Therefore, the inventors created a seed trait estimation model by combining image data of seeds captured with a general-purpose camera (RGB image) and data on seed traits such as the presence or absence of germination obtained by subjecting the seeds to germination testing, using this as training data. In other words, a seed trait estimation model was created using data that links seed images to seed trait data on a seed-by-seed basis as training data. However, it was found that when seed traits were estimated using the obtained seed trait estimation model and seed images, a problem arose in that the traits of the target seed could not be estimated.

[0005] Therefore, this disclosure provides a seed trait estimation device and a seed trait estimation method that can be used to estimate seed traits. [Means for solving the problem]

[0006] To achieve the above objective, the seed trait estimation device of this disclosure (hereinafter referred to as the "estimation device") includes a first image acquisition unit that acquires multi-wavelength images of one or more seeds, The system includes an estimation unit that estimates the characteristics of a seed by estimating the characteristics of one or more pixels constituting the seed in the multi-wavelength image using a first seed characteristic estimation model.

[0007] The seed trait estimation model generation device of this disclosure (hereinafter also referred to as the "generation device") includes a first learning unit that generates the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed trait data. The trait data is associated with one or more pixels that make up a seed in the multi-wavelength image.

[0008] The seed trait estimation method described herein (hereinafter referred to as the "estimation method") is a seed trait estimation method that is performed on a computer, A first acquisition step involves acquiring a multi-wavelength image in which one or more seeds are imaged, The method includes an estimation step of estimating the characteristics of a seed by estimating the characteristics of one or more pixels constituting the seed in the multi-wavelength image using a first seed characteristic estimation model.

[0009] The method for generating seed trait estimation models described herein (hereinafter referred to as the "generation method") is a method for generating seed trait estimation models that is performed on a computer, This includes a first learning step of generating the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with one or more pixels that make up a seed in the multi-wavelength image.

[0010] The program disclosed herein (hereinafter referred to as "the First Program") is installed on a computer. A first acquisition process to acquire a multi-wavelength image in which one or more seeds are captured, An estimation process is performed to estimate the characteristics of the seed by estimating the characteristics of one or more pixels constituting the seed in the multi-wavelength image using a first seed characteristic estimation model.

[0011] The program disclosed herein (hereinafter referred to as the "Third Program") is installed on a computer. This program executes a first learning process to generate a seed trait estimation model using machine learning, based on the first training data of the seed trait estimation model. The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with one or more pixels that make up a seed in the multi-wavelength image. [Effects of the Invention]

[0012] This disclosure provides a seed trait estimation device and a seed trait estimation method that can be used to estimate seed traits.

Brief Description of the Drawings

[0013] [Figure 1] FIG. 1 is a block diagram showing an example of an estimation system including an imaging device, a generation device, and an estimation device in Embodiment 1. [Figure 2] FIG. 2 is a block diagram showing an example of the hardware configuration of the generation device in Embodiment 1. [Figure 3] FIG. 3 is a block diagram showing an example of the hardware configuration of the estimation device in Embodiment 1. [Figure 4] FIG. 4 is a flowchart showing an example of the generation method, estimation method, and program in Embodiment 1. [Figure 5] FIG. 5 is a schematic diagram comparing the processing of the estimation device in Embodiment 1 with the treatment using a trait estimation model generated by associating the traits of seeds with the seed images. [Figure 6] FIG. 6 is a block diagram showing an example of an estimation system including an imaging device, a generation device, and an estimation device in Embodiment 2. [Figure 7] FIG. 7 is a flowchart showing an example of the generation method, estimation method, and program in Embodiment 2. [Figure 8] FIG. 8 is a block diagram showing an example of a sorting system including an imaging device, a generation device, an estimation device, and a sorting device in Embodiment 3. [Figure 9] FIG. 9 is a flowchart showing an example of the generation method, estimation method, sorting method, and program in Embodiment 3.

Embodiments of the Invention

[0014] <Definitions> In this specification, “multi-wavelength image” means an image containing two or more wavelength band components. The wavelength band components may also be, for example, the number of colors or bands. The number of wavelength band components contained in the multi-wavelength image may be two or more, and there is no particular upper limit. The number of wavelength band components contained in the multi-wavelength image may include the number necessary for generating a seed trait estimation model, for example, 1-500, 1-300, 3-500, 3-300, 4-500, 4-300, 5-500, or 5-300. The wavelength ranges of the wavelength band components are ultraviolet light (e.g., less than 400 nm), visible light (e.g., 400-750 nm), and / or infrared light (e.g., greater than 750 nm). The multi-wavelength image is, for example, an image acquired by an imaging device with a wavelength resolution (resolution of wavelength band components) of 10.5 nm or less, 10 nm or less, 8 nm or less, 6 nm or less, 5 nm or less, 2.5 nm or less, or 2 nm or less. The lower limit of the wavelength resolution of the imaging device is, for example, 1 nm or more or 2 nm or more. The wavelength resolution of the imaging device is, for example, 1 to 10.5 nm, 1 to 10 nm, 2 to 8 nm, 2 to 6 nm, 2 to 5 nm, 2 to 2.5 nm, or 2 to 2.3 nm. Examples of imaging devices having a wavelength resolution of 10.5 nm or less include hyperspectral cameras and multispectral cameras, which will be described later. Therefore, the multi-wavelength image may be, for example, a multispectral image or a hyperspectral image. The multi-wavelength image may exclude RGB images. In general, cameras that acquire RGB images have a wavelength resolution of about 30 nm.

[0015] In this specification, "seed" means an ovule of a seed plant that has matured after fertilization, and an ovule that has developed as is through parthenogenesis. The seed may be integrated with a fruit or other part. The seed includes, for example, vegetable seeds, flowering plant seeds, and grain seeds. The plants include, for example, Brassicaceae plants such as cabbage and Chinese cabbage; Cucurbitaceae plants such as melon, cucumber, and pumpkin; Solanaceae plants such as tomato, eggplant, and bell pepper; Apiaceae plants such as carrots; Amaryllidaceae plants such as leeks and onions; and Gentianaceae plants such as eustoma.

[0016] In this specification, the "seed trait" means a trait possessed by a seed that can be examined by a germination test, a biochemical test, or a pathological test. The seed trait can also be, for example, a trait that cannot be determined in the state of the seed. The "seed trait" can also be, for example, an inherent trait of the seed. Examples of the seed trait include the germination of the seed or the presence or absence of germination, the purity of the seed or the correctness of the purity, the disease of the seed (e.g., virus, bacterium, fungus, etc.) or the presence or absence of the disease of the seed, and the like.

[0017] Hereinafter, the present disclosure will be described in detail with reference to examples and the drawings. However, the present disclosure is not limited by the following description. In the following FIGS. 1 to 9, the same parts may be denoted by the same reference numerals and the description thereof may be omitted. In the drawings, for convenience of explanation, the structure of each part may be appropriately simplified, and the dimensional ratios and the like of each part may be different from the actual ones and may be shown schematically. Also, unless otherwise specified, the descriptions of each embodiment can be mutually referred to. In this specification, when the expression "~" is used, it is used in the sense of including the numerical values or physical values before and after it. In this specification, the expression "A and / or B" includes "only A", "only B", and "both A and B".

[0018] (Embodiment 1) This embodiment is an example of an estimation system including an imaging device, a generation device, and an estimation device of the present disclosure. FIG. 1 is a block diagram showing an estimation system 100 including an imaging device 1, a generation device 2, and an estimation device 3 of this embodiment. As shown in FIG. 1, the estimation system 100 of this embodiment includes an imaging device 1, a generation device 2, and an estimation device 3. The generation device 2 includes a second acquisition unit 21 and a first learning unit 22. The estimation device 3 includes a first acquisition unit 31 and an estimation unit 32. As shown in FIG. 1, the imaging device 1, the generation device 2, and the estimation device 3 can be connected one-way or two-way (communicable) via a communication line network 4 outside the estimation system 100.

[0019] The generation device 2 and estimation device 3 of this embodiment may be incorporated into a personal computer (PC) or a server as a system on which the program of this disclosure is installed. Alternatively, the generation device 2 and estimation device 3 of this embodiment may be a system consisting of one or more computers or servers capable of executing the program of this disclosure, i.e., a cloud computing system. The personal computers may constitute a computer cluster. Although not shown, the generation device 2 and estimation device 3 may be configured to connect to an external terminal of a system administrator via a communication network 4, allowing the system administrator to manage the generation device 2 and estimation device 3 from the external terminal. In this embodiment, the generation device 2 and estimation device 3 included in the estimation system 100 are each one, but there may be multiple instances of each.

[0020] The communication network 4 is not particularly restricted and can use any publicly known network, such as a wired or wireless connection. Examples of communication network 4 include the Internet, WWW (World Wide Web), telephone lines, LAN (Local Area Network), WiFi (Wireless Fidelity), etc.

[0021] Figure 2 illustrates a block diagram of the hardware configuration of the generation device 2. The generation device 2 includes, for example, a CPU (Central Processing Unit) 201, memory 202, bus 203, storage device 204, input device 206, display 207, and communication device (communication unit) 208, which are all arithmetic elements. Each part of the generation device 2 is connected via the bus 203 through its respective interface (I / F).

[0022] The CPU 201 operates in conjunction with other components via a controller (system controller, I / O controller, etc.) and is responsible for the overall control of the generation device 2. In the generation device 2, the CPU 201 executes the program (third program) 205 of this disclosure and other programs, and also reads and writes various types of information. Specifically, in this embodiment, the CPU 201 functions as the second acquisition unit 21 and the first learning unit. The generation device 2 includes a CPU as an arithmetic unit (arithmetic element), but may also include other arithmetic units such as a GPU (Graphics Processing Unit) or APU (Accelerated Processing Unit), or a combination of the CPU and these. The CPU 201 functions, for example, as one of the parts other than the storage unit in other embodiments.

[0023] Memory 202 includes main memory. This main memory is also called primary memory. When the CPU 201 performs processing, memory 202 reads various operational programs, such as the program 205 of this disclosure, which are stored in the storage device 204 (auxiliary storage device) described later. The CPU 201 then reads and decodes the data from memory 202 and executes the program. This main memory is RAM (Random Access Memory). Memory 202 may further include ROM (Read-Only Memory).

[0024] Bus 203 can also connect to external devices. Examples of such external devices include external storage devices (external databases, etc.) and printers. Estimation device 3 can connect to a communication network 4, for example, via a communication device 208 connected to the bus, and can also connect to the external devices via the communication network 4. Estimation device 3 can also connect to estimation terminal 2 via the communication device 208 and the communication network 4.

[0025] The storage device 204 is also called an auxiliary storage device, for example, in relation to the main memory. As described above, the storage device 204 stores an operating program including the program 205 of this disclosure. The storage device 204 includes, for example, a storage medium and a drive for reading from and writing to the storage medium. The storage medium is not particularly limited and may be internal or external, and examples include HD (hard disk), FD (floppy disk), CD-ROM, CD-R, CD-RW, MO, DVD, flash memory, memory card, etc., and the drive is not particularly limited. The storage device 204 may be, for example, a hard disk drive (HDD) in which the storage medium and the drive are integrated.

[0026] The generation device 2 further includes an input device 206 and an output device, a display 207. The input device 206 may include, for example, a pointing device such as a touch panel, trackpad, or mouse; a keyboard; imaging means such as a camera or scanner; a card reader such as an IC card reader or magnetic card reader; an audio input means such as a microphone; and so on. The display 207 may include, for example, an LED (light-emitting diode) display or a liquid crystal display. In this embodiment, the input device 206 and the display 207 are configured separately, but the input device 206 and the display 207 may be configured as an integrated unit, such as a touch panel display. Furthermore, in the generation device 2, the input device 206 and the display 207 are of any configuration and may be omitted, or one or more of them may be included.

[0027] Next, Figure 3 illustrates a block diagram of the hardware configuration of the estimation device 3. The estimation device 3 includes a CPU 301, memory 302, bus 303, storage device 304, input device (input unit) 306, and communication device (communication unit) 308. Each component of the estimation device 3 is connected via the bus 203 by its respective interface (I / F). In the estimation device 3, the CPU 301 functions as the first acquisition unit 31 and estimation unit 32, and the storage device 304 stores the program (first program) 305 of this disclosure. Except for these points, the description of each component of the estimation device 3 can be based on the description of each component of the generation device 2. In the estimation device 3, the storage device 304 may store the first estimation model described later.

[0028] Next, an example of the processing in the estimation system 100 of this embodiment will be explained using the flowchart in Figure 4 as an example. In this example, the generation device 2 generates a first seed trait estimation model (also called a "discriminator" or "learning model") based on multi-wavelength image data captured by the imaging device 1, and then the estimation device 3 performs the process of estimating the seed traits in the multi-wavelength image data using the multi-wavelength image data captured by the imaging device 1 and the first seed trait estimation model (first estimation model) generated by the generation device 2. In Figure 4, the imaging device 1 performs the processes S1 to S2, the generation device 2 performs the processes S3 to S7, and the estimation device 3 performs the processes S8 to S11.

[0029] Prior to processing in the generation device 2, in S1, multi-wavelength images (image data) to be used for generating the first estimation model are acquired using the imaging device 1 of the estimation system 100. Specifically, in S1, image data can be acquired by imaging one or more target seeds placed using the imaging device 1. The imaging device 1 can be an imaging device capable of acquiring two or more wavelength band components, and specific examples include multispectral cameras and hyperspectral cameras. For example, the following cameras can be used as multispectral cameras and hyperspectral cameras. (Multispectral camera) • msCAM: Manufactured by Spectral Corporation • VideometerLab4: Manufactured by Videometer Inc. • No product name: Manufactured by Fujifilm Corporation • SNAPSHOT SWIR: Made by imec • Color IR Monarch II: Manufactured by Unispectral Corporation (Hyperspectral camera) • PECIM FX Series: Manufactured by Spectral Imaging ·ULTRIS, FireflEYE: Manufactured by cubert • 4250VNIR, 4400SWIR, 4200M: Manufactured by HinaLea ·Pika NUV2, Pika L, Pika XC2, Pika NIR-320, Pika NIR-640: Manufactured by RESONON ·Hyspex Classic, Hyspex Baldur, Hyspex Mjolnir: Manufactured by Hyspex • AHS-003VIR: Manufactured by Abal Data Corporation

[0030] The multi-wavelength image is preferably included in the visible light region as its wavelength range, since this can improve the accuracy of the first estimation model. The wavelength ranges included in the multi-wavelength image are, for example, 450-800 nm, 450-1700 nm, 400-950 nm, 400-1000 nm, 350-1000 nm, 350-1700 nm, 330-800 nm, 330-2350 nm, or 330-2500 nm. The wavelength range of the multi-wavelength image can also be defined as the sum of the wavelength ranges of one or more imaging devices (the entire range).

[0031] In S1, the number of image data captured is limited to a number that can generate the first estimation model, and can be multiple, with no particular upper limit. The number of image data can be set to, for example, the number of image data containing 50 or more, 100 or more, 200 or more, 300 or more, 400 or more, or 500 or more target seeds. The number of pixels constituting the image data is limited to one or more, and can be determined by, for example, the imaging device 1, and can be 70,000 to 102 million pixels, 300,000 to 34.5 million pixels, or 900,000 to 12 million pixels.

[0032] In the aforementioned image data, for example, each wavelength band component is associated with a coordinate component and the signal intensity of the wavelength band component at each coordinate. Therefore, in the aforementioned image data, for one or more pixels, the coordinate data (coordinates: X, Y) in the two-dimensional plane (XY plane) of each pixel is associated with the wavelength data at the pixel with coordinates (X, Y), i.e., the wavelength band component (A) and its signal intensity value (B). Therefore, the aforementioned image data consists of X, Y, A, and B. Although the explanation has been given using the example of an image in a two-dimensional plane (XY plane), the aforementioned image data may also be a three-dimensional image in a three-dimensional plane (XYZ plane), etc. In this case, the aforementioned image data consists of, for example, X, Y, Z, A, and B.

[0033] In the following explanation, we will use the example of a two-dimensional plane in which the image data is acquired and the seed characteristics being germination or non-germination. However, the same procedure can be followed even if the space in which the image data is acquired is three-dimensional or more, and if the seed characteristics are other than germination or non-germination.

[0034] Then, in S2, the imaging device 1 transmits the obtained image data to the generation device 2 via the communication network 4 using a communication device (not shown). The imaging device 1 may transmit all of the image data obtained in S1, or it may transmit only a portion of it. The number of image data transmitted by the imaging device 1 should be equal to or greater than the number required to generate the first estimation model.

[0035] Next, processing in the generation device 2 is started. The generation device 2 uses training data (training data) that links the image data obtained by the imaging device 1 with the seed trait data (germination or non-germination) obtained by performing germination tests on the seeds in each image data, and generates a first estimation model by machine learning. In this embodiment, the generation device 2 uses data from one or more pixels of the image data, that is, data that links coordinate data, wavelength data, and seed trait data, as training data, and generates a first estimation model by machine learning.

[0036] First, in S3, the second acquisition unit 21 acquires image data transmitted from the imaging device 1 via the communication device 208 (second image acquisition). Next, in S4, the second acquisition unit 21 acquires seed trait data (germination or non-germination data) for each image data (second data acquisition). The seed trait data can be generated, for example, by performing an inspection on the seeds used to capture the image data, which is normally performed when determining the trait of the target to be estimated, and obtaining the inspection results. As a specific example, if the seed trait is seed germination or non-germination, a germination inspection is performed on each seed used in the imaging under normal conditions. Then, for each seed, the result of the germination inspection, i.e., the evaluation result of germination or non-germination, can be used as trait data for each seed.

[0037] In S5, the second acquisition unit 21 acquires (generates) training data by associating corresponding seed trait data with each image data (first data generation). Specifically, the second acquisition unit 21 generates training data (X, Y, A, B, C) with seed trait data by associating (tagging) the corresponding pixel trait data (C) with the data (X, Y, A, B) held by each pixel of one or more pixels in the image data. This association is performed on some or all of the one or more pixels of the image data, preferably on all of them. This association (annotation) may be performed automatically or manually. When the association is performed automatically, the seed trait data holds coordinate data (X, Y) to be associated with the trait data (C), and the association can be performed, for example, by associating the trait data (C) with the corresponding pixel based on the coordinate data. On the other hand, if the linking is performed manually, the linking can be carried out, for example, by the evaluator of the inspection inputting the evaluation of the traits for the image data using the input device 206, etc., and linking the trait data (C) to one or more pixels. In this embodiment, the second acquisition unit 21 acquires the image data and the seed trait data and links them to generate the learning data, but the second acquisition unit 21 may acquire pre-generated learning data. In this case, the learning data can be generated, for example, by linking the trait data to the image data using a computer outside the imaging device 1 or estimation system 100.

[0038] Next, in S6, the first learning unit 22 generates (learns) a first estimation model using machine learning with the learning data (first learning). The machine learning can be any machine learning method usable for classification, and specific examples include random forests, neural networks, support vector machines, principal component analysis, decision trees, etc. For example, when machine learning is performed using a neural network, the first learning unit 22 estimates the seed traits for one or more pixels of the input learning data. That is, the first learning unit 22 estimates the seed traits of one or more pixels of the input learning data using the parameters of the first estimation model. Then, the first learning unit 22 adjusts (updates) the parameters of the first estimation model using the error between the obtained estimation result and the trait data (C) associated with the input learning data so that this error is minimized. The parameter update can be performed as appropriate depending on the machine learning method. The first learning unit 22 repeatedly performs the estimation and updating until there are no more uninputted pixels in the learning data. This allows the first learning unit 22 to generate a first estimation model with parameters capable of estimating the seed characteristics of the pixels constituting each type of seed in a multi-wavelength image containing one or more seeds. The first learning unit 22 may adjust the parameters of the first estimation model using some of the pixels in the learning data, or it may adjust the parameters of the first estimation model using all of the pixels.

[0039] In the first learning unit 22, one or more pixels are input individually, that is, each pixel is input (1x1 pixel) and a first estimation model is generated by machine learning. However, the disclosure is not limited to this, and multiple pixels from one or more pixels may be input together and a first estimation model may be generated by machine learning. In this case, the first learning unit 22 divides the image data (multi-wavelength image) into multiple fractions, for example. Next, the first learning unit 22 calculates an average value for each data from the pixel data of each fraction and obtains the data for multiple pixels. Then, the first learning unit 22 inputs the data for multiple pixels and generates a first estimation model by machine learning. The size of each fraction can be, for example, 2x2 pixels, 3x3 pixels, 4x4 pixels, 5x5 pixels, 10x10 pixels, 50x50 pixels, or 100x100 pixels. Preferably, the sizes are 2x2 pixels, 3x3 pixels, 4x4 pixels, or 5x5 pixels, as these allow for the generation of a more accurate first estimation model.

[0040] The first learning unit 22 may use all or part of the wavelength data of the learning data for learning in generating the first estimation model. When part of the wavelength data is used for learning, the first learning unit 22 uses, for example, some wavelength band components and their signal intensities from the wavelength data for learning. When some wavelength band components are used, the lower limit of the number of wavelength band components is 2 or more, preferably 3 or more, 4 or more, or 5 or more, as this can further improve the accuracy of estimation, and more preferably 6 or more, as this can further improve the accuracy of estimation and suppress false negatives. The upper limit of the number of wavelength band components is not particularly limited and is determined according to the imageable wavelength range of the imaging device 1 and the interval of the measurement wavelengths. The upper limit of the wavelength band components is, for example, 2048 or less, 1024 or less, 512 or less, 256 or less, 128 or less, 64 or less, 50 or less, 32 or less, 21 or less, 20 or less, or 19 or less. The number of wavelength band components is, for example, 2-2048, 2-1024, 2-512, 3-256, 3-128, 3-128, 3-64, 4-50, 4-40, 4-30, 4-21, 4-20, 4-19, 5-21, 5-20, 5-19, 6-21, 6-20, or 6-19. When the number of wavelength band components is 3, the wavelength band components may exclude, for example, the combination of wavelength band components of an RGB image, i.e., the combination of 435.5nm, 546.1nm, and 700.0nm.

[0041] When using some wavelength band components in the generation of the first estimation model, the first learning unit 22 extracts some of the wavelength band components from the learning data. The extraction of the wavelength band components may be done randomly, or it may be done considering wavelength band components that are thought to contribute to the estimation of seed traits. As a specific example, the extraction of some wavelength band components can be done, for example, by extracting wavelength band components up to a predetermined rank (e.g., 2nd to 6th, 2nd to 5th, or 3rd) in order from the top of the feature importance ranking in the first estimation model.

[0042] Then, in S7, the generation device 2 transmits the first estimation model generated by the first learning unit 22 to the estimation device 3 via the communication network 4 using the communication device 208.

[0043] Next, processing in the estimation device 3 is started. The estimation device 3 uses the first estimation model generated by the generation device 2 to estimate the characteristics of the seeds in the image data transmitted from the imaging device 1. In this embodiment, the estimation device 3 estimates the characteristics of the seeds for one or more pixels of the image data, and based on the obtained estimation results, estimates the characteristics of each type of seed included in the image data.

[0044] First, in S8, the first acquisition unit 31 acquires the first estimation model transmitted from the generation device 2 via the communication device 308. Next, in S9, the first acquisition unit 31 acquires image data for estimating the characteristics of the seed (first image acquisition). In this embodiment, the seed characteristics are estimated using image data different from the image data used to generate the first estimation model, but this disclosure is not limited thereto. In this disclosure, the image data may be the image data used to generate the first estimation model, or it may be image data that is different from the image data used to generate the first estimation model in part or in whole. Also, if the first estimation model is generated using a part of the image data captured by the imaging device 1, the estimation device 3 may acquire the remainder of the image data as image data for estimating the characteristics of the seed. Furthermore, if image data different from the image data used to generate the first estimation model is used, the imaging device 1 newly captures a multi-wavelength image including the seed and transmits the obtained image data to the estimation device 3 via the communication network 4.

[0045] Next, in S10, the estimation unit 32 uses the first estimation model to estimate the characteristics of one or more pixels that constitute a seed in the image data. Specifically, the estimation unit 32 inputs wavelength data associated with one or more pixels that constitute a seed in the image data into the first model, that is, it inputs wavelength data associated with each pixel into the first model. Next, the estimation unit 32 uses the parameters of the first estimation model to estimate the characteristics of the seed to which the input pixel corresponds (for example, germination or non-germination). Then, the estimation unit 32 estimates the estimation result output from the first estimation model, that is, the characteristics of the seed, as the characteristics of the pixel. This estimation can also be called inference. The estimation unit 32 may estimate the characteristics of some of the pixels that constitute a seed in the image data, or it may estimate the characteristics of all of the pixels.

[0046] Next, in S11, the estimation unit 32 estimates the traits of each seed from the estimation results of the pixels constituting each seed. Based on the estimation results of the pixels, the estimation unit 32 may estimate the traits of each seed from the number of pixels associated with the estimation result that they have a predetermined trait or do not have a predetermined trait, or it may estimate the traits of each seed from the ratio of pixels associated with the estimation result that they have a predetermined trait or do not have a predetermined trait, based on the estimation results of the pixels. The ratio of pixels can be calculated, for example, as the ratio (Pt / P0) of the number of pixels (Pt) associated with the estimation result that they have a predetermined trait or do not have a predetermined trait to the total number of pixels (P0) of the seed to be estimated.

[0047] When the estimation unit 32 estimates the traits of each seed using the number of pixels, the estimation unit 32 estimates the traits of each seed as follows, as an example: If the number of pixels in the estimation result of the pixels constituting the seed to be estimated is 1 or more, 2 or more, or above a threshold, the estimation unit 32 estimates that the seed to be estimated has the predetermined trait. Specifically, if the number of pixels in the estimation result of the pixels constituting the seed to be estimated is 1 or more, 2 or more, or above a threshold, the estimation unit 32 estimates that the seed to be estimated has the trait of "germinating" or "not germinating," respectively. In addition to or instead of the above estimation, the estimation unit 32 may also estimate that the seed to be estimated has the trait of "not germinating" or "germinating," respectively, if the number of pixels in the estimation result of the pixels constituting the seed to be estimated is 0, 1 or less, or below a threshold.

[0048] Furthermore, when the estimation unit 32 estimates the traits of each seed using the proportion of the pixels, the estimation unit 32 estimates the traits of each seed as follows, as an example: If the proportion of pixels with a predetermined trait in the estimation result of the pixels constituting the seed to be estimated is greater than or equal to a threshold, the estimation unit 32 estimates that the seed to be estimated has the predetermined trait. Specifically, if the proportion of pixels with an estimation result of "germination" or "non-germination" in the estimation result of the pixels constituting the seed to be estimated is greater than or equal to a threshold, the estimation unit 32 estimates that the seed to be estimated has the trait of "germination" or "non-germination," respectively. In addition to or instead of the above estimation, the estimation unit 32 may also estimate that if the proportion of pixels with an estimation result of "germination" or "non-germination" in the estimation result of the pixels constituting the seed to be estimated is less than a threshold, the estimation unit 32 estimates that the seed to be estimated has the trait of "non-germination" or "germination," respectively.

[0049] The threshold can be set by evaluating the percentages of true positives, true negatives, false positives, and / or false negatives from the estimation results of the traits of each seed by the estimation unit 32 and the evaluation results from germination tests, etc., of each seed, so as to achieve the desired percentages of true positives, true negatives, false positives, and / or false negatives. The threshold may also be set using an ROC curve (Receiver Operating Characteristic curve). In this case, the threshold can be set, for example, by taking the percentage of false positives and the percentage of true positives as two axes, and setting the value on the ROC curve that achieves the shortest distance from the coordinate where the percentage of true positives is 100% and the percentage of false positives is 0%. Furthermore, when setting the threshold using the ROC curve, the threshold may also be set using a method that utilizes the Youden index. In addition, the threshold may also be set using AIC (Akaike Information Criterion) or a classification tree. The threshold may be set to a value specified by the user of the estimation system 100, such as 40%, 50%, 60%, 70%, 80%, or 90%.

[0050] Next, the estimation unit 32 repeatedly performs the same estimation until there are no more unestimated seeds. When there are no more seeds to be estimated by the estimation unit 32 in the image data, the estimation system 100 terminates processing.

[0051] In the estimation system 100 of this embodiment, which comprises an imaging device 1, a generation device 2, and an estimation device 3, as shown in Figure 5(A), multi-wavelength images are used, and for one or more pixels constituting seeds in the multi-wavelength image, seed characteristics are associated with the wavelength data to generate training data. The estimation system 100 then inputs the associated wavelength data and seed characteristic data for one or more pixels for each pixel, and generates a first estimation model by machine learning. Furthermore, in the estimation system 100 of this embodiment, the characteristics of each seed are estimated by estimating the characteristics of the seeds on a pixel-by-pixel basis for the multi-wavelength image containing the target seed, using the first estimation model obtained using the pixel-by-pixel training data. Therefore, according to the estimation system 100 of this embodiment, which comprises an imaging device 1, a generation device 2, and an estimation device 3, as shown in Figure 5(B), seed characteristics can be estimated more accurately than with an estimation model obtained using training data that associates seed characteristics with seed images.

[0052] In this embodiment, the generation device 2 that generates the first estimation model and the estimation device 3 that estimates seed traits using the first estimation model are configured as separate devices. However, the disclosure is not limited thereto, and the generation device 2 and the estimation device 3 may be configured as a single device. In this case, the disclosure provides, for example, a device that has the functions of each part of the generation device 2 and the estimation device 3. Furthermore, the first acquisition unit 31 of the generation device 2 and the second acquisition unit 21 of the estimation device 3 may be configured to perform the function of the other, with either one functioning independently.

[0053] In this embodiment, the estimation unit 32 estimates the characteristics of one or more pixels constituting a seed in the image data on a pixel-by-pixel basis using the first estimation model. However, the disclosure is not limited thereto, and the characteristics of multiple pixels in the one or more pixels may be estimated together. In this case, the estimation unit 32 divides the multi-wavelength image into multiple fractions, for example. Next, the estimation unit 32 calculates an average value for each data from the pixel data of each fraction to obtain the data for multiple pixels. Then, the estimation unit 32 estimates the characteristics of the multiple pixels using the first estimation model and the data for multiple pixels. Next, the estimation unit 32 estimates the characteristics of each seed from the estimation results of the multiple pixels constituting each seed. In this case, the estimation unit 32 can estimate the characteristics of each seed using the estimation results of the characteristics of multiple pixels instead of the estimation results of the characteristics on a pixel-by-pixel basis. The size of each pixel can be, for example, 2x2 pixels, 3x3 pixels, 4x4 pixels, 5x5 pixels, 10x10 pixels, 50x50 pixels, or 100x100 pixels. Preferably, each pixel is 2x2 pixels, 3x3 pixels, 4x4 pixels, or 5x5 pixels, as this allows for more accurate estimation.

[0054] The estimation device 3 of this embodiment estimates the characteristics of a seed by inputting the data of pixels constituting a seed from the image data. However, the disclosure is not limited thereto. Alternatively, the data of pixels that do not constitute a seed may be input from the multi-wavelength image data, i.e., both the data of pixels constituting a seed and the data of pixels that do not constitute a seed may be input, and the characteristics of the seed of each pixel may be estimated. Alternatively, pixel data may be input for all pixels of the image data, and the characteristics of the seed of each pixel may be estimated.

[0055] The estimation device 3 of this embodiment may include a detection unit for detecting one or more seeds in the image data. In this case, prior to the estimation processing by the estimation unit 32 (S10-S11), the estimation device 3 may use the detection unit to detect one or more seeds in the image data, and the estimation unit 32 may estimate the characteristics of the detected seeds, i.e., the pixels in the region detected as seeds. Seed detection can be performed, for example, by boundary extraction using binarization, extraction based on color difference from the background, edge detection, etc. By including the detection unit, the estimation device 3 can prevent false detection of seeds and reduce the number of pixels for which estimation processing is performed by the estimation unit 32, thereby shortening processing time and improving the accuracy of seed characteristic estimation.

[0056] The estimation device 3 of this embodiment may output the estimated results of the seed traits. In this case, the estimation device 3 may output to, for example, the display 307 which is the output device, an output device outside the estimation device 3, or an output device outside the estimation system 100. When the estimation device 3 outputs to the display 307, the display 307 displays the estimated results of the seed traits. Furthermore, the estimation device 3 may be configured to output or display, for example, the estimated results of the seed traits superimposed on a multi-wavelength image including the corresponding seed.

[0057] In this embodiment, a configuration capable of estimating seed germination or non-germination was described as an example, but the estimation system 100 may also estimate other seed traits. Furthermore, the combination of the plant from which the seed originates and the seed trait to be estimated can be arbitrarily set. For example, if the seed is from a Solanaceae plant or a Cucurbitaceae plant, the seed trait to be estimated could be, for example, whether or not the seed germinates, or whether or not it is diseased. If the seed is from a Brassicaceae plant, the seed trait to be estimated could be, for example, whether or not the genetic purity is correct.

[0058] In this embodiment, the estimation unit 32 estimated the characteristics of one or more pixels using only the first estimation model. However, the estimation unit 32 may also estimate the characteristics of one or more pixels by combining the first estimation model with other inference methods or statistical techniques. However, it is preferable to use the first estimation model alone as the method for estimating the characteristics of one or more pixels, in order to speed up the processing of the estimation device 3.

[0059] (Embodiment 2) This embodiment is another example of an estimation system including the imaging device, generation device, and estimation device of the present disclosure. Figure 6 is a block diagram showing an estimation system 100A comprising the imaging device 1, generation device 2A, and estimation device 3A of this embodiment. As shown in Figure 6, the generation device 2A of this embodiment includes an evaluation unit 23, an extraction unit 24, a second data generation unit 25, and a second learning unit 26, in addition to the configuration of the generation device 2 of Embodiment 1. Furthermore, the estimation device 3A of this embodiment uses a second seed trait estimation model (second estimation model) generated by the generation device 2A of Embodiment 2 instead of the first estimation model generated by the generation device 2 of Embodiment 1. The hardware configuration of the generation device 2A is the same as that of the generation device 2 of Figure 2, except that the CPU 201 has the configuration of the generation device 2A of Figure 6 instead of the configuration of the generation device 2 of Figure 2. Furthermore, the hardware configuration of the estimation device 3A is the same as that of the estimation device 3 in Figure 3, except that the CPU 301 is configured as shown in Figure 6 instead of the configuration of the estimation device 3 in Figure 3. Aside from these points, the configurations of the generation device 2A and estimation device 3A in Embodiment 2 are the same as those of the generation device 2 and estimation device 3 in Embodiment 1, and their explanations can be applied accordingly.

[0060] Next, an example of the processing of the estimation system 100A, which comprises the imaging device 1, generation device 2A, and estimation device 3A of this embodiment, will be explained using the flowchart in Figure 7.

[0061] First, the estimation system 100A of this embodiment acquires image data using the imaging device 1 in the same manner as the estimation system 100 of Embodiment 1 (S1). Then, the imaging device 1 transmits the obtained image data to the generation device 2 via the communication network 4 (S2).

[0062] Next, processing in the generation device 2A is started. In the generation device 2A, in the same manner as in the generation device 2, a first estimation model is generated by machine learning using training data that links the image data obtained by the imaging device 1 with the seed trait data (germination or non-germination). Then, the generation device 2A evaluates the degree of contribution of each wavelength band component to the estimation of the seed trait in the obtained first estimation model, and using the obtained evaluation results, a second estimation model is generated by machine learning using second training data that links some wavelength band components and their signal intensity in the image data with the seed trait data. In other words, in this embodiment, the generation device 2A uses the first estimation model to identify wavelength band components that contribute to the estimation of the target seed trait, and uses the identified wavelength band components to generate a second estimation model.

[0063] First, in step S3, the second acquisition unit 21 acquires image data transmitted from the imaging device 1 via the communication device 208, similar to the generation device 2 in Embodiment 1 (second image acquisition). Next, in step S4, the second acquisition unit 21 acquires seed trait data (germination or non-germination data) for each image data (second data acquisition). Then, in step S5, the second acquisition unit 21 acquires (generates) learning data by associating the corresponding seed trait data with each image data (first data generation).

[0064] Next, in step S6, similar to the generation apparatus 2 of Embodiment 1, the first learning unit 22 generates (learns) a first estimation model using machine learning with the learning data (first learning).

[0065] Next, in step S12, the evaluation unit 23 uses the first estimation model to evaluate the wavelengths, i.e., wavelength band components, in the image data that contribute to the estimation of the traits. Specifically, the evaluation unit 23 identifies the wavelength band components that contribute to the estimation of traits when the first estimation model estimates the traits of seeds. This identification can be determined, for example, according to the type of machine learning. If the first estimation model is generated using a random forest as the machine learning method, the identification can be performed based on the importance of the features in the random forest. That is, the identification can identify wavelength band components with high importance in the random forest as wavelength band components that relatively contribute to the estimation of traits. In this identification, the features may be ranked according to their importance, i.e., according to their importance.

[0066] As an example, the importance of the features may be evaluated using the degree of decrease in the recognition accuracy of the estimation model, referring to the references below. First, estimation models (1) and (2) below are generated using the training data. Next, the image data is estimated using (1) and (2) below, and the recognition accuracy of (1) and (2) below is calculated. Then, the difference in recognition accuracy between (1) and (2) is taken as the importance of the features, and features (wavelength band components) with a relatively large difference in recognition accuracy can be evaluated as more important features. (1) Estimation model trained using the features of the training data as they are (2) An estimation model is trained using modified training data, in which one of the features of the training data is shuffled within the training data, while the remaining values ​​are left unchanged, and the modified training data is used as the training data. Reference: L. Breiman, “Random Forests”, Machine Learning, 45(1), 5-32, 2001.

[0067] The importance of the aforementioned features may be evaluated using the Gini impurity described in the aforementioned references.

[0068] In S13, the extraction unit 24 uses the evaluation results from the evaluation unit 23 to extract spectral data of two or more predetermined wavelengths (wavelength band components) from the image data acquired by the second acquisition unit 21 as training data. The lower limit of the number of wavelength band components extracted by the extraction unit 24 is two or more, preferably three or more, four or more, or five or more, in order to further improve the accuracy of estimation, and more preferably six or more, in order to further improve the accuracy of estimation and suppress false negatives. The upper limit of the number of wavelength band components is not particularly limited and is determined according to the imageable wavelength range and wavelength resolution of the imaging device 1. For example, the upper limit of the number of wavelength band components is 2048 or less, 1024 or less, 512 or less, 256 or less, 128 or less, 64 or less, 50 or less, 32 or less, 21 or less, 20 or less, or 19 or less. The number of wavelength band components is, for example, 2-2048, 2-1024, 2-512, 3-256, 3-128, 3-128, 3-64, 4-50, 4-40, 4-30, 4-21, 4-20, 4-19, 5-21, 5-20, 5-19, 6-21, 6-20, or 6-19. The extracted wavelength band components are preferably those selected in order from the highest importance in the ranking of feature quantities in the first estimation model. In this embodiment, the extraction unit 24 performs extraction from image data used to generate the first estimation model, but the disclosure is not limited thereto, and the extraction unit 24 may perform extraction from image data that is partially or entirely different from the image data used to generate the first estimation model.

[0069] In S14, the second data generation unit 25 associates seed trait data with one or more pixels constituting seeds in the training data extracted by the extraction unit 24, thereby generating second training data. The generation of the second training data by the second data generation unit 25 can be carried out in the same manner as the first data generation in the estimation system 100 of Embodiment 1, and the explanation therefor can be applied.

[0070] In S15, the second learning unit 26 generates a second seed trait determination model (second estimation model) by machine learning using the second learning data (second learning). The machine learning in the second learning unit 26 can be performed in the same way as the machine learning in S6. The machine learning in S15 is the same as the machine learning used to generate the first estimation model. Furthermore, it is preferable that the second learning unit 26 generates the second estimation model using all wavelength band components included in the second learning data. This allows the second learning unit 26 to generate a second estimation model with parameters capable of estimating the seed traits of pixels constituting each seed for a multi-wavelength image containing one or more seeds. In this embodiment, the machine learning in S6 and S15 is the same, but may be different.

[0071] Then, in S16, the generation device 2A transmits the second estimation model generated by the second learning unit 26 to the estimation device 3A via the communication network 4 using the communication device 208.

[0072] Next, processing in the estimation device 3A is started. The estimation device 3A uses the second estimation model generated by the generation device 2A to estimate the characteristics of the seeds in the image data transmitted from the imaging device 1. In this embodiment, the estimation device 3A uses the second estimation model instead of the first estimation model in the estimation device 3 of Embodiment 1 to estimate the characteristics of the seeds for one or more pixels of the image data, and estimates the characteristics of each type of seed included in the image data based on the obtained estimation results. Except for this point, the estimation device 3A of Embodiment 2 estimates the characteristics of the target seed in the same manner as the estimation device 3 of Embodiment 1.

[0073] First, in S17, the first acquisition unit 31 acquires the second estimation model transmitted from the generation device 2A via the communication device 308. Next, in S9, the first acquisition unit 31 acquires image data for estimating the characteristics of the seed (first image acquisition). In this embodiment, the seed characteristics are estimated using image data different from the image data used to generate the first and second estimation models, but this disclosure is not limited thereto. In this disclosure, the image data may be the image data used to generate the first and second estimation models, or it may be image data that is different in part or all from the image data used to generate the first and second estimation models. Furthermore, if the first and second estimation models are generated using a portion of the image data captured by the imaging device 1, the estimation device 3A may acquire the remainder of the image data as image data for estimating the characteristics of the seed. Furthermore, if different image data is used from the image data used to generate the first and second estimation models, the imaging device 1 captures new multi-wavelength images including seeds and transmits the obtained image data to the estimation device 3 via the communication network 4.

[0074] Next, in S10A, the estimation unit 32 estimates the characteristics of one or more pixels constituting a seed in the image data using the second estimation model. Then, in S11A, the estimation unit 32 estimates the characteristics of each type of seed from the estimation results of the pixels constituting each type of seed. S10A and S11A can be carried out in the same manner as in S10 and S11 of Embodiment 1, except that the second estimation model is used instead of the first estimation model. In this disclosure, the estimation unit 32 inputs all wavelength band components of the image data and estimates the characteristics of one or more pixels, but the disclosure is not limited thereto, and the characteristics of one or more pixels may be estimated by inputting only a portion of the wavelength band components of the image data. In this case, the portion of the wavelength band components includes some or all of the wavelength band components used to generate the second estimation model, preferably including all of them.

[0075] Next, the estimation unit 32 repeatedly performs the same estimation until there are no more unestimated seeds. When there are no more seeds to be estimated by the estimation unit 32 in the image data, the estimation system 100A terminates processing.

[0076] In the estimation system 100A of this embodiment, which includes an imaging device 1, a generation device 2A, and an estimation device 3A, the wavelength band components contributing to the estimation of seed traits are identified in the first estimation model, and a second estimation model is generated using these wavelength band components. Then, the estimation system 100A estimates the seed traits using the obtained second estimation model. In the estimation system 100 of this embodiment, data for all wavelength band components is not required for trait estimation, and the amount of input data can be reduced, thereby improving processing speed while maintaining estimation accuracy.

[0077] In this embodiment, the generation device 2A that generates the second estimation model and the estimation device 3A that estimates seed traits using the second estimation model are configured as separate devices. However, this disclosure is not limited thereto, and the generation device 2A and the estimation device 3A may be configured as a single device. In this case, for example, this disclosure provides a device that has the functions of each part of the generation device 2A and the estimation device 3A. Furthermore, the first acquisition unit 31 of the generation device 2A and the second acquisition unit 21 of the estimation device 3A may be configured to consist of either one and perform the function of the other.

[0078] (Embodiment 3) This embodiment is an example of a sorting system including the imaging device, generation device, estimation device, and sorting device of the present disclosure. Figure 8 is a block diagram of a sorting system 200 comprising the imaging device 1, generation device 2, estimation device 3, and sorting device 5 of this embodiment. As shown in Figure 8, the sorting system 200 of this embodiment includes a sorting device (sorting unit) 5 in addition to the configuration of the estimation system 100 of Embodiment 1. Furthermore, the sorting device 5 can be connected (communicated) in one or both directions via a communication network 4 outside the sorting system 200 through the imaging device 1, generation device 2, and estimation device 3.

[0079] Next, an example of the processing of the sorting system 200, which comprises the imaging device 1, generation device 2, estimation device 3, and sorting device 5 of this embodiment, will be explained using the flowchart in Figure 9.

[0080] First, steps S1 to S11 are carried out in the same manner as in Embodiment 1, with respect to the imaging device 1, generation device 2, and estimation device 3. Next, in S18, the estimation device 3 transmits the estimation results of the obtained seed traits to the sorting device 5 via the communication network 4. Preferably, the estimation results include coordinate data of the seed to be estimated.

[0081] Next, processing by the sorting device 5 is started. The sorting device 5 uses the seed trait estimation results obtained from the estimation device 3 to select seeds that have the desired traits.

[0082] In S19, the sorting device 5 acquires the estimation results transmitted from the generating device 2 via a communication device (not shown). Next, in S20, the sorting device 5 sorts the seeds based on the seed traits, i.e., the estimated seed traits. Specifically, the sorting device 5 can sort seeds having the desired traits by recovering seeds estimated to have the desired traits, or recovering or removing seeds estimated not to have the desired traits, based on the estimation results. The sorting device 5 can use a sorting machine that can be used for seed sorting, and a specific example is the Sorter (manufactured by Seed-X Technologies).

[0083] The sorting device 5 repeatedly performs the same seed sorting until there are no more seeds that meet the estimation criteria. When there are no more seeds that the estimation unit 32 has estimated in the image data, the sorting system 200 terminates processing.

[0084] In the sorting system 200 of this embodiment, which includes an imaging device 1, a generation device 2, an estimation device 3, and a sorting device 5, the sorting device 5 can sort seeds that are estimated to have the desired traits using the seed trait estimation results obtained by the estimation device 3. Therefore, according to the sorting system 200 of this embodiment, it is possible to sort seeds having predetermined traits without performing germination tests or the like.

[0085] (Embodiment 4) The program of this embodiment is a program that causes a computer to execute the estimation method, generation method, and / or selection method described herein. In the program of this embodiment, “procedure” can also be said to be, for example, “process” or “instruction.” The program of this embodiment may also be recorded on, for example, a computer-readable storage medium. The storage medium is, for example, a non-transitory computer-readable storage medium. The storage medium is not particularly limited and includes, for example, random access memory (RAM), read-only memory (ROM), hard disk (HD), optical disk, floppy disk (FD), and the like. [Examples]

[0086] Next, embodiments of the present disclosure will be described. However, the present disclosure is not limited to the following embodiments.

[0087] [Example 1] We have confirmed that the estimation device described in this disclosure can estimate whether tomato seeds have germinated or not.

[0088] (1) Generation of the first estimation model An estimation model for estimating traits related to tomato seed germination was generated using the generation device 2 of the estimation system 100 of Embodiment 1, following the procedure below. First, tomato seeds of Solanaceae plants (one medium-sized tomato variety (Fruitica), two large-sized tomato varieties (Momotaro Peace, Momotaro York), manufactured by Takii Seed Co., Ltd.) were prepared. The prepared seeds were imaged using a multi-wavelength camera Pika L (manufactured by RESONON, wavelength range: 400-1000 nm, wavelength resolution: 3.7 nm, measurement wavelength interval: approximately 2 nm) and Pika NIR-640 (manufactured by RESONON, wavelength range: 900-1700 nm, wavelength resolution: 5.3 nm, measurement wavelength interval: approximately 2.5 nm), and image data including brightness information (wavelength band components (approximately 600 wavelengths in total for approximately 400-1700 nm) and their signal intensity values) was acquired for each image capture. The number of seeds from which image data was acquired was 500 or more. Next, the tomato seeds used for imaging were subjected to a germination test to evaluate the germination and non-germination of each seed. Then, labels indicating germination (0) or non-germination (1), presence or absence of seeds, and seed number were associated with each pixel of the image data to generate first training data for machine learning. Using this first training data, a first estimation model capable of estimating the classification of germination and non-germination was generated by machine learning (random forest, hereafter the same).

[0089] (2) Estimation of tomato seed germination Image data was acquired for tomato seeds not used in Example 1(1) (one medium-sized tomato variety (Fruitica) and two large-sized tomato varieties (Momotaro Peace, Momotaro York)) in the same manner as in Example 1(1). Then, for the germination / non-germination traits of each seed, the traits of each pixel were estimated using the image data and the first estimation model generated in Example 1(1), and then the traits of the seed were estimated from the traits of each pixel. In estimating the traits of the seed, a threshold was set in advance based on the proportion of pixels with traits estimated to indicate germination to the total number of pixels. If the threshold was exceeded, the seed to be estimated was considered a seed that would germinate. After the estimation, seeds estimated to have the germination trait (n=100 for each group) were selected, and germination tests were conducted to calculate the germination rate (germination vigor) over a certain period. In addition, germination tests were conducted on unselected seeds (n=100) from the same lot to calculate the germination rate. These results are shown in Table 1 below.

[0090] [Table 1]

[0091] Table 1 above shows the germination rate results. As shown in Table 1, the germination rate of the sorted tomato seeds was improved compared to the unsorted tomato seeds. From these results, it was found that the germination and non-germination of tomato seeds can be estimated using the first estimation model described above.

[0092] (3) Generation of the second estimation model According to the first estimation model generated in Example 1(1) above, the germination and non-germination traits of tomato seeds can be estimated. Therefore, using the generation device 2A in the estimation system 100A of Embodiment 2, wavelength band components important for trait estimation were extracted and used in the first estimation model to generate a second estimation model, and it was confirmed whether the germination and non-germination of tomato seeds could be estimated using the second estimation model.

[0093] In the first estimation model described above, the wavelength band components were ranked from highest to lowest in order of feature importance. The top 2, 3, 4, 5, or 6 wavelength band components were then identified. Next, tomato seeds (one medium-sized tomato variety (Frutica) and two large-sized tomato varieties (Momotaro Peace, Momotaro York)) were prepared. The prepared seeds were imaged using a multi-wavelength camera to acquire image data. The top 2, 3, 4, 5, or 6 wavelength band components were extracted from the image data to generate training data. Then, the training data was labeled with seed characteristics to generate training data (second training data). An estimation model (second estimation model) was generated by machine learning in the same manner as in Example 1(1), except that the second training data was used instead of the first training data.

[0094] (4) Estimation of tomato seed germination Image data was obtained for tomato seeds not used in Example 1(3) (one medium-sized tomato variety (Fruitica) and two large-sized tomato varieties (Momotaro Peace, Momotaro York)). Then, for the germination / non-germination traits of each seed, the traits of each pixel were estimated using the image data and the second estimation model generated in Example 1(3), and then the seed traits were estimated from the traits of each pixel. The second estimation model was input by extracting and inputting data of the wavelength band components used to generate the second estimation model from the image data. In estimating the seed traits, a threshold was set in advance based on the ratio of pixels with traits estimated to germinate to the total number of pixels, and if it was above the threshold, the seed to be estimated was considered a seed that would germinate. After the estimation, seeds estimated to have germination / non-germination traits (n=100 for each group) were selected, and germination tests were conducted to calculate the germination rate. Furthermore, germination tests were conducted on unsorted seeds from the same lot (n=100), and the germination rate (germination vigor) over a certain period was calculated. These results are shown in Table 2 below.

[0095] [Table 2]

[0096] Table 2 shows the germination rate results. As shown in Table 2, the germination rate of the sorted tomato seeds was improved compared to the unsorted tomato seeds. It was found that using three or more wavelength band components during training ensures high accuracy in germination rate, and that by using at least the top three wavelengths, seed sorting can be performed by the second estimation model with the same efficiency as the first estimation model. Furthermore, it was found that using four or more wavelengths allows for more accurate estimation of germination and non-germination. In addition, when using eight varieties including Momotaro Peace and Momotaro York, results showing a similar trend to that of Fruitica were obtained. From these results, it was found that the germination and non-germination of tomato seeds can be estimated by the second estimation model. Furthermore, in estimating the seed characteristics of each pixel using the second estimation model described above, each image data was divided into units of 2x2 pixels, 3x3 pixels, 4x4 pixels, 5x5 pixels, 10x10 pixels, 50x50 pixels, or 100x100 pixels, and the seed characteristics of each fraction (data of multiple pixels) were estimated. Then, the germination or non-germination of tomato seeds was estimated from the seed characteristics of the multiple pixels. As a result, the seed characteristics were estimated well regardless of which multiple pixels were used to estimate the seed characteristics. It was also found that when estimating the seed characteristics of multiple pixels, reducing the size of the fraction, especially to 5x5 pixels or less, allowed for more accurate estimation of germination or non-germination.

[0097] Although the data is not shown, the germination and non-germination of tomato seeds were similarly estimated when the first and second estimation models were generated using luminance information obtained using only Pika L. Similarly, the germination and non-germination of tomato seeds were also estimated when the first and second estimation models were generated using luminance information obtained using the hyperspectral camera AHS-003VIR (manufactured by AVAL DATA, wavelength range: 450~1700nm, wavelength resolution 10.5nm). From these results, it was concluded that the luminance information in the visible light range acquired by either camera is important for estimating seed germination and non-germination.

[0098] Furthermore, in generating the first and second estimation models, instead of inputting 1x1 pixels, each image data was divided into 2x2, 3x3, 4x4, 5x5, 10x10, 50x50, or 100x100 pixel units, and the average value of multiple pixels in each fraction was input. The second estimation model was generated in the same manner as before. Then, data was input pixel by pixel for one or more pixels of the image data into the obtained second estimation model, and the seed characteristics of each pixel were estimated. As a result, the seed characteristics were estimated well regardless of the size of the multiple pixel data used to generate the first and second estimation models. It was also found that when generating the first and second estimation models using the multiple pixel data, reducing the size of the fractions, especially to 5x5 pixels or less, resulted in first and second estimation models that could estimate germination and non-germination with greater accuracy.

[0099] [Example 2] We have confirmed that the estimation device described in this disclosure can estimate whether tomato seeds have germinated or not.

[0100] A comparative study was conducted with a commercially available tomato seed sorting machine. In estimation device 3 of Example 2, the second estimation model generated using the three wavelength band components in Example 1(3) was used. For unsorted tomato seeds (Frutica, test group C), the germination and non-germination of the seeds were estimated in the same manner as in Example 1(4), except that the second estimation model generated using the three wavelength band components was used, and each was collected (test group A: estimated to have germinated, test group B: estimated to have not germinated). In addition, a commercially available seed color sorter was used to select tomato seeds that had germinated or not. The sorting machine sorts based on the intensity of the seed coat color (RGB) (test group D: group targeting darker colors, test group E: group targeting lighter colors). Then, for test groups A to E (n=100 in each group), the germination rate (germination vigor), final germination rate (germination rate), and final non-germination rate (non-germination) within a certain period were calculated. These results are shown in Table 3.

[0101] [Table 3]

[0102] Table 3 shows the results for germination rate (germination vigor), final germination rate (germination rate), and final non-germination rate (non-germination). As shown in Table 3, when using a commercially available color sorter, seeds targeted with darker colors by the color sorter (test group D) showed a slightly improved germination rate compared to unsorted seeds (test group C), but seeds targeted with lighter colors by the color sorter (test group E) showed almost no difference compared to unsorted seeds (test group C). Furthermore, among the seeds estimated to germinate according to the second estimation model (test group A), almost all germinated, and the germination rate was significantly improved compared to unsorted seeds (test group C) and seeds sorted using a color sorter (test groups D and E). On the other hand, among the seeds estimated not to germinate (test group B), the germination rate decreased and the non-germination rate increased compared to unsorted seeds (test group C).

[0103] Next, we investigated whether the germination and non-germination of seeds could be estimated using the second estimation model from Example 2 and multi-wavelength images (image data) of seeds from eight varieties of tomato plants. Specifically, we estimated the germination and non-germination of each seed in the same manner as before, except that we used multi-wavelength images (image data) of each variety instead of multi-wavelength images (image data) of variety A, and collected each group (n=100). Then, we conducted germination tests on each group and calculated the final germination rate and final non-germination rate (non-germination). As a result, in the seeds estimated to have germinated, the germination rate improved significantly and the non-germination rate decreased, as in Table 3 above. Also, in the seeds estimated to have not germinated, the non-germination rate improved, as in Table 3 above.

[0104] These results indicate that sorting using the estimation results obtained with the estimation device described herein can achieve superior accuracy compared to sorting using commercially available color sorters. Furthermore, it was estimated that estimation models generated using seeds of different varieties of the same plant can also be used to estimate traits.

[0105] [Example 3] We have confirmed that the estimation device described herein can estimate the germination / non-germination of melon seeds, onion seeds, and eustoma seeds.

[0106] A first estimation model was generated in the same manner as in Example 1(1), except that melon seeds (Lennon Star, manufactured by Takii Seed Co., Ltd.), a member of the Cucurbitaceae family, were used instead of the aforementioned tomato seeds. Next, a second estimation model was generated in the same manner as in Examples 1(2) and (3), except that the aforementioned melon seeds were used instead of the aforementioned tomato seeds, using three wavelengths as the wavelength band components to be used as training data. This model was then used to estimate the germination and non-germination of the melon seeds. Then, germination tests were conducted on the seeds estimated to have the traits of germination and non-germination, and on unsorted seeds (n=50 in each group), and the germination rate (germination vigor) within a certain period was calculated. As a result, the unsorted seeds had a germination rate of 66% and a non-germination rate of 34%, while the seeds estimated to have germinated had a germination rate of 100% and a non-germination rate of 0%. Furthermore, among the seeds estimated to be unsuccessful, the germination rate was 52% and the unsuccessful germination rate was 48%. These results show that the estimation device described herein can also estimate the germination and unsuccessful germination of melon seeds.

[0107] Furthermore, a first estimation model was generated in the same manner as in Example 1(1), except that onion seeds of the Amaryllidaceae family (Neo Earth, manufactured by Takii Seed Co., Ltd.) were used instead of the tomato seeds. Next, a second estimation model was generated in the same manner as in Examples 1(2) and (3), except that onion seeds were used instead of the tomato seeds, with three wavelengths used as the wavelength band components for training data, and this model was used to estimate the germination and non-germination of onion seeds. Then, germination tests were conducted on the seeds estimated to have the traits of germination and non-germination, and on unsorted seeds (n=100 in each group), and the germination rate (germination vigor) within a certain period was calculated. As a result, the germination rate for unsorted seeds was 69% and the non-germination rate was 31%, while the germination rate for seeds estimated to have germinated was 90% and the non-germination rate was 10%. Furthermore, among the seeds estimated to be unsuccessful, the germination rate was 61% and the unsuccessful germination rate was 39%. These results show that the estimation device described herein can also estimate the germination and unsuccessful germination of onion seeds.

[0108] Furthermore, a first estimation model was generated in the same manner as in Example 1(1), except that Eustoma seeds (F1 folder cherry, manufactured by Takii Seed Co., Ltd.), a plant of the Gentianaceae family, were used instead of the tomato seeds. Next, a second estimation model was generated in the same manner as in Examples 1(2) and (3), except that Eustoma seeds were used instead of the tomato seeds, with three wavelengths used as the wavelength band components for training data. This model was then used to estimate the germination and non-germination of Eustoma seeds. Then, germination tests were conducted on the seeds estimated to have the traits of germination and non-germination, and on unsorted seeds (n=50 in each group), and the germination rate (germination vigor) within a certain period was calculated. As a result, the unsorted seeds had a germination rate of 86% and a non-germination rate of 14%, while the seeds estimated to have germinated had a germination rate of 93% and a non-germination rate of 7%. Furthermore, among the seeds estimated to be unsuccessful, the germination rate was 66% and the unsuccessful germination rate was 34%. These results show that the estimation device described herein can also estimate the germination and unsuccessful germination of Eustoma seeds.

[0109] Furthermore, for the melon seeds, onion seeds, and eustoma seeds, in estimating the seed characteristics of each pixel using the second estimation model, the image data was divided into units of 2x2 pixels, 3x3 pixels, 4x4 pixels, 5x5 pixels, 10x10 pixels, 50x50 pixels, or 100x100 pixels, and the seed characteristics of each fraction (data of multiple pixels) were estimated. Then, the germination or non-germination of the melon seeds, onion seeds, and eustoma seeds was estimated from the seed characteristics of the multiple pixels. As a result, the seed characteristics were estimated well regardless of which multiple pixels were used to estimate the seed characteristics. It was also found that when estimating the seed characteristics of multiple pixels, reducing the size of the fraction, especially to 5x5 pixels or less, allowed for more accurate estimation of germination or non-germination.

[0110] Furthermore, in generating the first and second estimation models, instead of inputting 1x1 pixels, each image data was divided into 2x2, 3x3, 4x4, 5x5, 10x10, 50x50, or 100x100 pixel units, and the average value of multiple pixels in each fraction was input. The second estimation model was generated in the same manner as before. Then, data was input pixel by pixel for one or more pixels of the image data into the obtained second estimation model, and the seed characteristics of each pixel were estimated. As a result, the seed characteristics were estimated well regardless of the size of the multiple pixel data used to generate the first and second estimation models. It was also found that when generating the first and second estimation models using the multiple pixel data, reducing the size of the fractions, especially to 5x5 pixels or less, resulted in first and second estimation models that could estimate germination and non-germination with greater accuracy.

[0111] [Example 4] We have confirmed that the estimation device described in this disclosure can estimate the genetic purity of cabbage seeds.

[0112] (1) Generation of the first estimation model An estimation model for estimating traits related to the genetic purity of cabbage seeds was generated using the generation device 2 of the estimation system 100 of Embodiment 1, following the procedure below. First, cabbage seeds of the Brassicaceae family (YR Harusora, manufactured by Takii Seed Co., Ltd.) were prepared. A first estimation model was generated in the same manner as in Embodiment 1(1), except that cabbage seeds were used instead of tomato seeds, and the estimation target was genetic purity instead of the presence or absence of germination. Then, a second estimation model was generated in the same manner as in Embodiments 1(2) and (3), except that cabbage seeds were used instead of tomato seeds, and the estimation target was genetic purity instead of the presence or absence of germination, using three wavelengths as the wavelength band components to be used as training data, and the genetic purity of the cabbage seeds was estimated using this model. The number of seeds from which image data was acquired was 500 or more. Next, the cabbage seeds used for imaging were subjected to biochemical testing to evaluate the genetic purity (F1 purity) of each offspring. Specifically, after extracting genomic DNA from the cabbage seeds, a PCR reaction was performed using a primer set capable of specifically amplifying regions that can identify a particular variety, along with the genomic DNA, to obtain amplified fragments. The genetic purity was then evaluated based on the difference in length (genotype) of the obtained amplified fragments. A label indicating the correctness of the genetic purity was then associated with each pixel of the image data to generate first training data for machine learning. Using this first training data, a first estimation model capable of estimating the correctness of the genetic purity classification was generated using machine learning (random forest).

[0113] (2) Estimation of the genetic purity of cabbage seeds Image data was acquired for cabbage seeds not used in Example 4(1) in the same manner as in Example 4(1). Then, for the genetic purity trait of each seed, the trait of each pixel was estimated using the image data and the first estimation model generated in Example 4(1), and then the trait of the seed was estimated from the trait of each pixel. In estimating the trait of the seed, a threshold was set in advance based on the proportion of pixels with the trait estimated to have correct genetic purity to the total number of pixels. If the threshold was exceeded, the seed to be estimated was considered to have the trait of correct genetic purity. After the estimation, seeds estimated to have the trait of correct genetic purity (n=100) were selected, and the biochemical tests were performed to calculate the proportion of seeds with the trait of correct genetic purity. In addition, the biochemical tests were performed on unselected seeds (n=100) from the same lot, and the proportion of seeds with the trait of correct genetic purity was calculated. As a result, 93% of unselected seeds possessed traits that indicated correct genetic purity, whereas 97% of selected seeds possessed traits that indicated correct genetic purity, indicating an increase in the proportion of seeds with correctly identified traits. These results demonstrate that the first estimation model can estimate the genetic purity of cabbage seeds. Furthermore, for the cabbage seeds, in estimating the traits of each pixel using the first estimation model, the image data was divided into 2x2, 3x3, 4x4, 5x5, 10x10, 50x50, or 100x100 pixel units, and the traits of the seeds in each fraction (data with multiple pixels) were estimated. The genetic purity of the cabbage seeds was then estimated from the traits of the multiple pixels. As a result, the traits of the seeds were estimated well regardless of the size of the multiple pixels used for seed trait estimation. Furthermore, it was found that when estimating the traits of seeds with multiple pixels, reducing the size of the fractions, especially to 5x5 pixels or less, allows for more accurate estimation of genetic purity.

[0114] Furthermore, in generating the first estimation model, instead of inputting 1x1 pixels, each image data was divided into 2x2, 3x3, 4x4, 5x5, 10x10, 50x50, or 100x100 pixel units, and the average value of multiple pixels in each fraction was input. The first estimation model was generated in the same manner as before. Then, data was input pixel by pixel for one or more pixels of the image data into the obtained first estimation model, and the seed characteristics of each pixel were estimated. As a result, the seed characteristics were estimated well regardless of the size of the multiple pixel data used to generate the first estimation model. It was also found that when generating the first estimation model using the multiple pixel data, reducing the size of the fractions, especially to 5x5 pixels or less, resulted in a first estimation model that could estimate the genetic purity of cabbage seeds with greater accuracy.

[0115] [Example 5] We have confirmed that the estimation device described in this disclosure can estimate whether or not tomato seeds are affected by disease.

[0116] (1) Generation of the first estimation model An estimation model for estimating the presence or absence of disease in tomato seeds was generated using the generation device 2 of the estimation system 100 of Embodiment 1, following the procedure below. First, tomato seeds (TTM-171, manufactured by Takii Seed Co., Ltd.) were prepared. A first estimation model was generated in the same manner as in Embodiment 1(1), except that TTM-171 tomato seeds were used as the tomato seeds and the estimation target was the presence or absence of disease instead of the presence or absence of germination. Then, a second estimation model was generated in the same manner as in Embodiments 1(2) and (3), except that TTM-171 tomato seeds were used as the tomato seeds and the estimation target was the presence or absence of disease instead of the presence or absence of germination, using three wavelengths as the wavelength band components to be used as training data, and this was used to estimate the presence or absence of disease in the tomato seeds. The number of seeds from which image data was acquired was 500 or more. Next, the tomato seeds used for imaging were subjected to pathological examination to evaluate the presence or absence of fungal growth in each seed. Then, a label indicating the presence or absence of mold was associated with each pixel of the image data, generating first training data for machine learning. Using this first training data, a first estimation model capable of estimating the classification of whether or not mold is present was generated using machine learning (random forest).

[0117] (2) Estimation of the presence or absence of diseases in tomato seeds Image data was acquired for tomato seeds not used in Example 5(1) in the same manner as in Example 5(1). Then, the presence or absence of mold growth in each seed was estimated using the image data and the first estimation model generated in Example 5(1), and the characteristics of each pixel were estimated from the characteristics of each pixel. In estimating the characteristics of the seeds, a threshold was set in advance based on the proportion of pixels with characteristics estimated to indicate mold growth to the total number of pixels. If the threshold was exceeded, the seeds to be estimated were considered to be seeds that would develop mold. After the estimation, seeds estimated to develop mold (n=100) were selected, and the pathological examination was performed to calculate the proportion of seeds that would develop mold. The proportion of seeds that would develop mold when estimated to be free of mold growth was also calculated. Regarding the threshold, a threshold was predetermined based on the proportion of pixels with traits estimated to be free from mold growth to the total number of pixels. If the threshold was above the threshold, the seeds to be estimated were estimated to be free from mold growth; if the threshold was below the threshold, the seeds to be estimated were estimated to be free from mold growth. After the estimation, the seeds estimated to be free from mold growth (n=100) were selected, and the pathological examination was performed to calculate the proportion of seeds that would develop mold. Furthermore, the same pathological examination was performed on unselected seeds from the same lot (n=100) to calculate the proportion of seeds that would develop mold. A similar experiment was conducted on four different lots of tomato seeds. These results are shown in Table 4 below.

[0118] [Table 4]

[0119] As shown in Table 4 above, the proportion of mold growth decreased in seeds estimated to be free of mold compared to the unselected seed group. In seeds estimated to be affected by mold, the proportion of mold growth increased compared to the unselected seed group. These results show that the presence or absence of mold growth in tomato seeds can be estimated using the first estimation model. Furthermore, for the tomato seeds, in estimating the seed characteristics of each pixel using the first estimation model, the image data was divided into units of 2x2 pixels, 3x3 pixels, 4x4 pixels, 5x5 pixels, 10x10 pixels, 50x50 pixels, or 100x100 pixels, and the seed characteristics of each fraction (data of multiple pixels) were estimated. Then, the presence or absence of mold growth in the tomato seeds was estimated from the seed characteristics of the multiple pixels. As a result, the seed characteristics were estimated well regardless of the size of the multiple pixels used for seed characteristic estimation. Furthermore, it was found that when estimating the characteristics of seeds with multiple pixels, reducing the size of the fractions, especially to 5x5 pixels or less, allows for more accurate estimation of the presence or absence of mold growth.

[0120] Furthermore, in generating the first estimation model, instead of inputting 1x1 pixels, each image data was divided into 2x2, 3x3, 4x4, 5x5, 10x10, 50x50, or 100x100 pixel units, and the average value of multiple pixels in each fraction was input. The first estimation model was generated in the same manner as before. Then, data was input pixel by pixel for one or more pixels of the image data into the obtained first estimation model, and the characteristics of the seed for each pixel were estimated. As a result, the characteristics of the seed were estimated well regardless of the size of the multiple pixel data used to generate the first estimation model. It was also found that when generating the first estimation model using the multiple pixel data, reducing the size of the fraction, especially to 5x5 pixels or less, resulted in a first estimation model that could estimate the presence or absence of mold on tomato seeds with greater accuracy.

[0121] While the present disclosure has been described above with reference to embodiments and examples, the present disclosure is not limited to the above embodiments and examples. Various modifications to the structure and details of the present disclosure are possible, as can be understood by those skilled in the art within the scope of the present disclosure.

[0122] <Note> Some or all of the above embodiments and examples may be described as follows, but are not limited to the following. <Seed trait estimation device> (Note 1) A first image acquisition unit that acquires a multi-wavelength image in which one or more seeds are captured, A seed trait estimation device, comprising: an estimation unit that estimates the traits of a seed by estimating the traits of one or more pixels constituting the seed in the multi-wavelength image using a first seed trait estimation model. (Note 2) The seed trait estimation device according to Appendix 1, wherein the estimation unit inputs wavelength data associated with one or more pixels constituting a seed in the multi-wavelength image into the first seed trait estimation model and estimates the output trait as the trait of the pixel. (Note 3) In the estimation unit, In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels that are estimated to have a predetermined trait is greater than or equal to a threshold, and / or In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait is less than a threshold, A seed trait estimation device according to Appendix 1 or 2, which estimates that the aforementioned seed possesses a desired trait. (Note 4) Includes a detection unit that detects one or more seeds in the multi-wavelength image, The seed trait estimation device according to any one of the appendices 1 to 3, wherein the estimation unit estimates the traits of the detected seeds. (Note 5) The first seed trait estimation model described above is a learned model trained using machine learning with training data for seed trait estimation models. The aforementioned training data is The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with one or more pixels that constitute a seed in the multi-wavelength image. A seed trait estimation device as described in any of the appendices 1 to 4. (Note 6) It includes a first learning unit that generates the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. A seed trait estimation device according to any one of the appendices 1 to 5, wherein the trait data is associated with one or more pixels constituting seeds in the multi-wavelength image. (Note 7) A second image acquisition unit that acquires multi-wavelength images of one or more seeds, A second data acquisition unit acquires seed morphological data in the multi-wavelength image, The seed trait estimation device according to Appendix 6, comprising: a first data generation unit that associates the trait data with one or more pixels constituting seeds in the multi-wavelength image and generates training data. (Note 8) An evaluation unit that uses the first seed trait estimation model to evaluate the wavelengths contributing to the estimation of the traits in the multi-wavelength image, An extraction unit that uses the obtained evaluation results to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data, A second data generation unit generates second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, A seed trait estimation device according to any one of the appendices 1 to 7, comprising a second learning unit that generates a second seed trait estimation model by machine learning using the second training data. (Note 9) The seed trait estimation device described in Appendix 8, wherein the estimation unit uses the second seed trait estimation model instead of the first seed trait estimation model. (Note 10) The seed trait estimation device according to any one of the appendices 1 to 9, wherein the seed traits are seed germination or the presence or absence of seed germination, seed purity or the correctness of seed purity, and / or seed disease or the presence or absence of seed disease. (Note 11) The trait estimation device is a server, which is a seed trait estimation device as described in any of the appendices 1 to 10. <Seed sorting device> (Note 12) A trait estimation unit that estimates the traits of a seed from a multi-wavelength image in which one or more seeds are captured, It includes a sorting unit that sorts the seeds based on the aforementioned traits, The trait estimation unit includes a seed trait estimation device described in any of the appendices 1 to 11, and is a seed sorting device. <Generator for generating seed trait estimation models> (Note 13) It includes a first learning unit that generates the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed trait data. A device for generating seed trait estimation models, wherein the trait data is associated with one or more pixels constituting seeds in the multi-wavelength image. (Note 14) A second image acquisition unit that acquires multi-wavelength images of one or more seeds, A second data acquisition unit acquires seed morphological data in the multi-wavelength image, A seed trait estimation model generation device according to Appendix 13, comprising: a first data generation unit that associates the trait data with one or more pixels constituting seeds in the multi-wavelength image to generate training data. (Note 15) An evaluation unit that uses the first seed trait estimation model to evaluate the wavelengths contributing to the estimation of the traits in the multi-wavelength image, An extraction unit that uses the obtained evaluation results to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data, A second data generation unit generates second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, A seed trait estimation model generation device according to Appendix 13 or 14, comprising a second learning unit that generates a second seed trait estimation model by machine learning using the second learning data. <Method for estimating seed traits> (Note 16) A method for estimating seed traits performed on a computer, A first acquisition step involves acquiring a multi-wavelength image in which one or more seeds are imaged, A seed trait estimation method, comprising an estimation step of estimating the traits of a seed by estimating the traits of one or more pixels constituting the seed in the multi-wavelength image using a first seed trait estimation model. (Note 17) The seed trait estimation method according to Appendix 16, wherein in the estimation step, the estimation unit inputs wavelength data associated with one or more pixels constituting a seed in the multi-wavelength image into the first seed trait estimation model and estimates the output trait as the trait of the pixel. (Note 18) In the estimation process described above, In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels with estimation results for a predetermined trait are greater than or equal to a threshold, and / or In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait is less than a threshold, A method for estimating seed traits, as described in Appendix 16 or 17, which estimates that the aforementioned seeds possess a desired trait. (Note 19) The process includes a detection step of detecting one or more seeds in the multi-wavelength image, The seed trait estimation method described in any of appendices 16 to 18, wherein the estimation step involves estimating the traits of the detected seeds. (Note 20) The first seed trait estimation model described above is a learned model trained by machine learning using training data for seed trait estimation models. The aforementioned training data is The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with one or more pixels that constitute a seed in the multi-wavelength image. A method for estimating seed traits as described in any of the appendices 16 to 19. (Note 21) This includes a first learning step of generating the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. A seed trait estimation method according to any one of the appendices 16 to 20, wherein the trait data is associated with one or more pixels constituting the seeds in the multi-wavelength image. (Note 22) A second image acquisition step involves acquiring a multi-wavelength image of one or more seeds, A second data acquisition step for acquiring seed phenotypic data in the multi-wavelength image, The seed trait estimation method according to Appendix 21, comprising a first data generation step of associating the trait data with one or more pixels constituting seeds in the multi-wavelength image to generate training data. (Note 23) An evaluation step is performed to evaluate the wavelengths that contribute to the estimation of the traits in the multi-wavelength image using the first seed trait estimation model described above. Using the obtained evaluation results, an extraction step is performed to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data. A second data generation step involves generating second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, A method for estimating seed traits according to any one of the appendices 16 to 22, comprising a second learning step of generating a second seed trait estimation model by machine learning using the second learning data. (Note 24) The seed trait estimation method described in Appendix 23, wherein in the estimation step, the second seed trait estimation model is used instead of the first seed trait estimation model. (Note 25) The seed trait estimation method described in any of Appendix 16 to 24, wherein the seed trait is seed germination or the presence or absence of seed germination, seed purity or the correctness of seed purity, and / or seed disease or the presence or absence of seed disease. <Seed Selection Method> (Note 26) A method for selecting seeds that is performed on a computer, A trait estimation step in which the traits of a seed are estimated from a multi-wavelength image in which one or more seeds are captured, The process includes a selection step of selecting the seeds based on the aforementioned traits, The method wherein the trait estimation step is performed by the seed trait estimation method described in any of appendices 16 to 25. <Method for generating seed trait estimation models> (Note 27) A method for generating a seed trait estimation model that runs on a computer, This includes a first learning step of generating the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. A method for generating a seed trait estimation model, wherein the trait data is associated with one or more pixels constituting the seeds in the multi-wavelength image. (Note 28) A second image acquisition step involves acquiring a multi-wavelength image of one or more seeds, A second data acquisition step for acquiring seed phenotypic data in the multi-wavelength image, A method for generating a seed trait estimation model as described in Appendix 27, comprising a first data generation step of associating the trait data with one or more pixels constituting seeds in the multi-wavelength image to generate training data. (Note 29) An evaluation step is performed to evaluate the wavelengths that contribute to the estimation of the traits in the multi-wavelength image using the first seed trait estimation model described above. Using the obtained evaluation results, an extraction step is performed to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data. A second data generation step involves generating second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, A method for generating a seed trait estimation model according to Appendix 27 or 28, comprising a second learning step of generating a second seed trait estimation model by machine learning using the second learning data. <Program 1> (Note 30) On the computer, A first acquisition process to acquire a multi-wavelength image in which one or more seeds are captured, A program that performs an estimation process to estimate the characteristics of a seed by estimating the characteristics of one or more pixels constituting the seed in the multi-wavelength image using a first seed characteristic estimation model. (Note 31) The estimation process, in which the estimation unit inputs wavelength data associated with one or more pixels constituting a seed in the multi-wavelength image into the first seed trait estimation model and estimates the output trait as the trait of the pixel, is the program as described in Appendix 30. (Note 32) In the estimation process described above, In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels with estimation results for a predetermined trait are greater than or equal to a threshold, and / or In the estimation results of the traits of one or more pixels constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait is less than a threshold, A program according to appendix 30 or 31, which presumes that the aforementioned seeds possess a desired trait. (Note 33) The process includes detecting one or more seeds in the multi-wavelength image, The estimation process described above involves a program, as described in any of appendices 30 to 32, that estimates the characteristics of the detected seeds. (Note 34) The first seed trait estimation model described above is a learned model trained by machine learning using training data for seed trait estimation models. The aforementioned training data is The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with one or more pixels that constitute a seed in the multi-wavelength image. The program described in any of the appendices 30 to 33. (Note 35) This includes a first learning process that generates the seed trait estimation model by machine learning using first training data of the seed trait estimation model, The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. A program according to any one of the appendices 30 to 34, wherein the trait data is associated with one or more pixels constituting seeds in the multi-wavelength image. (Note 36) A second image acquisition process to obtain a multi-wavelength image capturing one or more seeds, A second data acquisition process for acquiring seed morphological data in the multi-wavelength image, The program described in Appendix 35, which includes a first data generation process that associates the trait data with one or more pixels constituting seeds in the multi-wavelength image to generate training data. (Note 37) An evaluation process is performed using the first seed trait estimation model to evaluate the wavelengths that contribute to the estimation of the traits in the multi-wavelength image, Using the obtained evaluation results, an extraction process is performed to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data. A second data generation process generates second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, A program according to any one of the appendices 30 to 36, which includes a second learning process for generating a second seed trait estimation model by machine learning using the second learning data. (Note 38) The program described in Appendix 37 uses the second seed trait estimation model instead of the first seed trait estimation model in the estimation process. (Note 39) The aforementioned seed characteristics are seed germination or the presence or absence of seed germination, seed purity or the correctness of seed purity, and / or seed disease or the presence or absence of seed disease, as described in any of the appendices 30 to 38. <Program 2> (Note 40) On the computer, A trait estimation process that estimates the traits of a seed from a multi-wavelength image in which one or more seeds are captured, This includes a sorting process for selecting the seeds based on the aforementioned traits, The aforementioned trait estimation process is performed by a program that is executed by any of the processes described in appendices 30 to 39. <Third Program> (Note 41) On the computer, This program executes a first learning process to generate a seed trait estimation model using machine learning, based on the first training data of the seed trait estimation model. The first training data mentioned above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. A program in which the trait data is associated with one or more pixels that constitute a seed in the multi-wavelength image. (Note 42) A second image acquisition process to obtain a multi-wavelength image capturing one or more seeds, A second data acquisition process for acquiring seed morphological data in the multi-wavelength image, The program described in Appendix 41, which includes a first data generation process that associates the trait data with one or more pixels constituting seeds in the multi-wavelength image to generate training data. (Note 43) An evaluation process is performed using the first seed trait estimation model to evaluate the wavelengths that contribute to the estimation of the traits in the multi-wavelength image, Using the obtained evaluation results, an extraction process is performed to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data. A second data generation process generates second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, The program described in Appendix 41 or 42, which includes a second learning process for generating a second seed trait estimation model by machine learning using the second training data. <Recording medium> (Note 44) A computer-readable recording medium containing a program described in any of the appendices 30 to 43. [Industrial applicability]

[0123] As described above, this disclosure provides a seed trait estimation device and a seed trait estimation method that can be used to estimate seed traits. Furthermore, this disclosure allows for the appropriate sorting of seeds using the obtained determination results. For this reason, this disclosure is extremely useful, for example, in the seed and seedling field, quality control, and the like. [Explanation of Symbols]

[0124] 1. Imaging device 2, 2A generator 21 Second acquisition section 22 First Learning Section 23 Evaluation Department 24 Extraction part 25 Second Data Generation Unit 26. Second Learning Section 3, 3A estimation device 31 First acquisition section 32 Estimation part 4. Communication Network 5. Sorting device

Claims

1. A first image acquisition unit that acquires a multi-wavelength image in which one or more seeds are captured, The system includes an estimation unit that estimates the characteristics of a seed by estimating the characteristics of each pixel constituting the seed in the multi-wavelength image, using a first seed characteristic estimation model, which is a learning model trained by machine learning using training data of a seed characteristic estimation model. The aforementioned training data is The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with each pixel that constitutes a seed in the multi-wavelength image. In the estimation unit, In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated to have a predetermined trait is greater than or equal to a threshold, and / or In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait are less than a threshold, It is presumed that the aforementioned seeds possess the desired traits, The estimation of the aforementioned seed traits is for predicting whether the seeds will germinate or not, and / or for estimating the genetic purity of the seeds. The aforementioned multi-wavelength image is a hyperspectral image or a multispectral image, in a seed trait estimation device.

2. Includes a detection unit that detects one or more seeds in the multi-wavelength image, The seed trait estimation device according to claim 1, wherein the estimation unit estimates the traits of the detected seeds.

3. An evaluation unit that uses the first seed trait estimation model to evaluate the wavelengths contributing to the estimation of the traits in the multi-wavelength image, An extraction unit that uses the obtained evaluation results to extract spectral data of two or more predetermined wavelengths from multi-wavelength images of one or more seeds as training data, A second data generation unit generates second training data by associating seed trait data with one or more pixels constituting seeds in the aforementioned training data, The seed trait estimation device according to claim 1, further comprising a second learning unit that generates a second seed trait estimation model by machine learning using the second learning data.

4. The seed trait estimation device according to claim 3, wherein the estimation unit uses the second seed trait estimation model instead of the first seed trait estimation model.

5. A first learning unit generates the seed trait estimation model by machine learning using first training data of the seed trait estimation model, A second image acquisition unit that acquires multi-wavelength images of one or more seeds, A second data acquisition unit acquires seed morphological data in the multi-wavelength image, The system includes a first data generation unit that associates the trait data with each pixel constituting a seed in the multi-wavelength image and generates training data, The first training data described above is: The system includes multi-wavelength images of one or more seeds and seed phenotype data. The trait data is associated with each pixel that constitutes a seed in the multi-wavelength image. The seed trait estimation device according to claim 1.

6. A method for estimating seed traits performed on a computer, A first acquisition step involves acquiring a multi-wavelength image in which one or more seeds are imaged, The process includes an estimation step of estimating the characteristics of a seed by estimating the characteristics of each pixel constituting the seed in the multi-wavelength image, using a first seed characteristic estimation model, which is a learning model trained by machine learning using training data of a seed characteristic estimation model. The training data includes multi-wavelength images of one or more seeds and seed trait data. The trait data is associated with each pixel that constitutes a seed in the multi-wavelength image. In the estimation process described above, In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated to have a predetermined trait is greater than or equal to a threshold, and / or In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait are less than a threshold, It is presumed that the aforementioned seeds possess the desired traits, The estimation of the aforementioned seed traits is for predicting whether the seeds will germinate or not, and / or for estimating the genetic purity of the seeds. A method for estimating seed traits, wherein the multi-wavelength image is a hyperspectral image or a multispectral image.

7. On the computer, A first acquisition process to acquire a multi-wavelength image in which one or more seeds are captured, An estimation process is performed to estimate the characteristics of the seed by using a first seed characteristic estimation model, which is a learning model trained by machine learning using training data of a seed characteristic estimation model, to estimate the characteristics of each pixel constituting the seed in the multi-wavelength image. The training data includes multi-wavelength images of one or more seeds and seed trait data. The trait data is associated with each pixel that constitutes a seed in the multi-wavelength image. In the estimation process described above, In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated to have a predetermined trait is greater than or equal to a threshold, and / or In the estimation results of the traits of each pixel constituting the seed, if the number and / or proportion of pixels that are estimated not to have a predetermined trait are less than a threshold, The seed is presumed to possess the desired traits, and the estimation of the seed's traits is a prediction of whether the seed will germinate or not, and / or an estimation of the seed's genetic purity. The aforementioned multi-wavelength image is a hyperspectral image or a multispectral image. program.

Images