Discrimination device, information processing device, discrimination processing program, teacher data generation method, discrimination method, and generation method.

Abstract

To allow even a beginner to differentiate a double blossom.SOLUTION: A learned model is generated by a machine learning that uses teacher data including a learning TIF image, first to third learning ranks, and correct data. A differentiation device (10) includes: a rank determination unit (11) for determining a first target rank to a third target rank for each of a plurality of target seedlings (502); and a determination unit (12) for inputting the target TIF image and the first to third target ranks input into the learned model and determining whether a specific target seedling (502) is a double blossom seedling.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a classification device, an information processing device, a classification processing program, training data, a classification method, and a generation method. [Background technology]

[0002] Stocks as cut flowers have been widely cultivated and are one of the most popular flowers among consumers. Stocks are flowers of the Brassicaceae family that bloom from October to March and are annual plants that grow to a height of about 20 cm to about 80 cm. Stocks come in single and double flower varieties, with the double flower varieties having a more voluminous feel and higher commercial value than the single flower varieties. Therefore, in order to make a high profit from stocks, it is important to accurately distinguish between single and double flower seedlings when the stocks are in the seedling stage. Hereinafter, distinguishing between single and double flower seedlings of stocks will be referred to as "double flower identification."

[0003] However, when visually identifying double-ply seeds, the worker must rely on his or her senses to distinguish subtle differences between the cotyledons of each seedling, which requires skill and time. Therefore, research has been conducted on techniques that allow even beginners to easily identify double-ply seeds. A "beginner" in double-ply identification includes a person who has no experience in double-ply identification, or a worker with less than 36 months of experience.

[0004] For example, non-patent document 1 discloses a method for distinguishing between double-flowered seedlings by obtaining features of cotyledons from images of germinated stock seedlings by image processing, and then using these features to optimize the boundary value between double-flowered seedlings and single-flowered seedlings. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Choi Won-hwan, Dohi Makoto, Ishizuka Nobuaki, "Development of a robot for precise management of flower and vegetable seedlings (Part 1) - Identification of stock seedlings by image processing -", Journal of the Japanese Society of Agricultural Machinery 66(2), pp.68-75, 2004 Summary of the Invention [Problem to be solved by the invention]

[0006] In the method disclosed in Non-Patent Document 1, morphological features and color features are acquired as features. The morphological features are acquired by dividing them into multiple items according to the morphology of the cotyledons (simple or cotyledons). In addition, in the above-mentioned method, in order to determine the boundary value between double-flowered seedlings and single-flowered seedlings, which serves as the discrimination criterion, the average value and standard deviation are calculated for each feature, and the feature is converted into a relative value. In other words, in the above-mentioned method, the average value and standard deviation must be calculated for each feature and item, and then a process of converting the features into relative values ​​must be performed. Therefore, the pre-processing stage before double-flowered seedling discrimination is already time-consuming.

[0007] In addition, in the method disclosed in Non-Patent Document 1, after the first stage of double-fold discrimination is performed for each item of morphological features, the first stage discrimination results of each item and color features are comprehensively evaluated to perform the second stage of double-fold discrimination (final double-fold discrimination). Therefore, the double-fold discrimination itself becomes complicated. For these reasons, it cannot be said that even beginners can easily discriminate double-fold using the above method.

[0008] One aspect of the present invention has been made in consideration of the above problems, and an object of the present invention is to enable even beginners to easily distinguish between double-fold ovaries. [Means for solving the problem]

[0009] In order to solve the above-mentioned problems, an identification device according to one embodiment of the present invention is an identification device for identifying double-flowered seedlings from seedlings of a plurality of target stocks before identifying the double-flowered seedlings, and a trained model is generated by machine learning using teacher data including: (i) learning light information of reflected light reflected by the seedlings of a plurality of learning stocks after it has been determined whether or not the seedlings are double-flowered seedlings; (ii) a learning rank determined by ranking the learning attribute values ​​of the cotyledons among the seedlings of the plurality of learning stocks; and (iii) correct answer data indicating whether or not the seedlings are double-flowered seedlings. The identification device is equipped with: a ranking determination unit that obtains, for each of the seedlings of the plurality of target stocks, a target attribute value of the cotyledons from target light information of the reflected light reflected by the seedlings, and ranks the target attribute values ​​among the seedlings of the plurality of target stocks to determine the target rank; and a determination unit that determines whether or not a seedling of a specific target stock is the double-flowered seedling by inputting the target light information and the target rank for a specific seedling of the target stock into the trained model.

[0010] According to the above configuration, by simply inputting the target light information and the target ranking into the trained model, the judgment unit can obtain the judgment result of whether or not a specific target stock seedling is a double-flowered seedling as an output result from the trained model. The target light information can be obtained from the reflected light reflected by the target stock seedling, and the target ranking can be obtained by ranking the target attribute values ​​obtained from each of the multiple target light information.

[0011] Here, even beginners can easily input the target light information and the like into the trained model, obtain the target light information, and rank the target attribute values. For example, a captured image captured by a general imaging device such as a digital camera may be used as the target light information. Since a general imaging device such as a digital camera can be easily operated even by beginners, the captured image as the target light information can be easily obtained. For these reasons, even beginners can easily perform double classification.

[0012] Next, even if the data is based on target light information obtained from seedlings of the same target stock, the target attribute value varies depending on where the seedlings of the target stock are placed. In this regard, according to the above configuration, the target ranking is input to the trained model instead of the target attribute value. Since the target ranking is data in which the target attribute value of each of the seedlings of multiple target stocks is converted into a relative value, the use of the target ranking can eliminate the influence of the placement locations of the seedlings of multiple target stocks on the target attribute value. Therefore, the degree to which the identification accuracy of the identification device varies depending on the placement locations of the seedlings of multiple target stocks can be reduced.

[0013] In the identification device according to one aspect of the present invention, the learning rankings included in the teacher data include (i) a first learning ranking determined by ranking the learning hue angles of the cotyledons among the seedlings of the plurality of learning stocks, (ii) a second learning ranking determined by ranking the learning saturation of the cotyledons among the seedlings of the plurality of learning stocks, and (iii) a third learning ranking determined by ranking the learning leaf area of ​​the cotyledons among the seedlings of the plurality of learning stocks, and the ranking determination unit determines, from the target light information, the target hue angle and the target saturation of the cotyledons as the target attribute values ​​for each of the seedlings of the plurality of target stocks. The determination unit may acquire a double-flowered seedling from a specific one of the target stock seedlings, (α) rank the target hue angle among the multiple target stock seedlings to determine a first target rank, (β) rank the target saturation among the multiple target stock seedlings to determine a second target rank, and (γ) rank the target leaf area among the multiple target stock seedlings to determine a third target rank, and the determination unit may determine whether the specific target stock seedling is a double-flowered seedling by inputting the target light information, the first target rank, the second target rank, and the third target rank for the specific target stock seedling into the trained model.

[0014] According to the above configuration, the identification accuracy of eight-fold identification can be improved compared to when the input data to the trained model is target light information and any one of the first to third target orders, and when the input data is target light information and any two of the first to third target orders.

[0015] In an appraisal device according to one embodiment of the present invention, the training light information included in the teacher data may include first training light information of first reflected light of artificial light reflected by the seedlings of the training stock, and second training light information of second reflected light of natural light reflected by the seedlings of the training stock.

[0016] According to the above configuration, the teacher data includes the first and second learning light information. Therefore, the trained model can correctly determine whether a seedling is a double-flowered seedling with a certain level of accuracy, even if the seedlings of the multiple target stocks from which the target light information and target ranking are obtained are irradiated with either artificial light or natural light. Here, artificial light is light emitted from artificial light sources such as light bulbs, fluorescent lights, and LEDs. Moreover, natural light is light originating from nature such as sunlight, moonlight, and reflected light from clouds, as well as all light other than light emitted from artificial light sources. Therefore, even if the seedlings of the multiple target stocks are irradiated with either artificial light or natural light, even beginners can easily distinguish whether they are double-flowered while maintaining a certain level of discrimination accuracy.

[0017] In one embodiment of the present invention, the appraisal device further includes a display control unit that controls a display device that displays target images representing seedlings of the multiple target stocks on a display screen, and the display control unit may highlight either a first image portion representing the seedlings of the target stock that have been determined by the judgment unit to be double-flowered seedlings, or a second image portion representing the seedlings of the target stock that have been determined by the judgment unit to be single-flowered seedlings, among the target images displayed on the display screen.

[0018] According to the above configuration, the judgment result of the judgment unit can be grasped simply by looking at the display screen of the display device, thereby improving the convenience of double-fold discrimination using the discrimination device.

[0019] In order to solve the above-mentioned problems, an information processing device according to an aspect of the present invention includes the discrimination device according to any one of the aspects of the present invention. According to the above-mentioned configuration, the same effects as those of the discrimination device according to the aspect of the present invention are achieved.

[0020] In order to solve the above problem, the training data according to one embodiment of the present invention is training data for generating a trained model that is applied to a classification device that performs machine learning to classify double-flowered seedlings from the seedlings of the multiple target stocks, and includes a ranking unit that acquires target attribute values ​​of cotyledons from target light information of the light reflected by the seedlings for each of the seedlings of the multiple target stocks before classifying double-flowered seedlings, and ranks the target attribute values ​​among the seedlings of the multiple target stocks to determine the target rank, and includes training light information of the light reflected by the seedlings of the multiple training stocks after it has been determined whether the seedlings are double-flowered seedlings or not, a training rank determined by ranking the training attribute values ​​of the cotyledons among the seedlings of the multiple training stocks, and correct answer data indicating whether the seedlings are double-flowered seedlings or not, and is used in a process in which the classification device determines whether the seedlings of the specific target stock are double-flowered seedlings by inputting the target light information and the target rank of a specific target stock seedling into the trained model. According to the above configuration, the same effect as that of the classification device according to one embodiment of the present invention can be achieved.

[0021] In order to solve the above-mentioned problems, one embodiment of the present invention provides an identification method for identifying double-flowered seedlings from seedlings of a plurality of target stocks before identifying the double-flowered seedlings, wherein a trained model is generated by machine learning using teacher data including: (i) learning light information of reflected light reflected by the seedlings of a plurality of learning stocks after it has been determined whether the seedlings are double-flowered seedlings; (ii) a learning rank determined by ranking the learning attribute values ​​of the cotyledons among the seedlings of the plurality of learning stocks; and (iii) correct answer data indicating whether the seedlings are double-flowered seedlings. The method includes a target rank determination step for obtaining, for each of the seedlings of the plurality of target stocks, a target attribute value of the cotyledons from the target light information of the reflected light reflected by the seedlings, and ranking the target attribute values ​​among the seedlings of the plurality of target stocks to determine a target rank; and a determination step for inputting the target light information and the target rank for a specific seedling of the target stock into the trained model to determine whether the seedling of a specific target stock is the double-flowered seedling. According to the above configuration, the same effects as those of the discrimination device according to one aspect of the present invention are achieved.

[0022] In order to solve the above problem, a generating method according to one aspect of the present invention is a generating method of a trained model used in a discrimination device for discriminating double-flowered seedlings from seedlings of a plurality of target stocks before the discrimination of the double-flowered seedlings, the generating method including: a data acquiring step of acquiring (i) training light information of reflected light from the seedlings, (ii) training attribute values ​​of the cotyledons, and (iii) correct answer data representing whether the seedlings are double-flowered seedlings or not, for the seedlings of the plurality of training stocks after it has been determined whether the seedlings are double-flowered seedlings or not, a training rank determining step of determining a training rank for each of the seedlings of the plurality of training stocks by ranking the training attribute values ​​among the seedlings of the plurality of training stocks, and a generating step of inputting a data set including the training light information, the correct answer data, and the training rank into a learning model as teacher data, and machine learning the learning model to generate the trained model. According to the above configuration, the same effect as that of the discrimination device according to one aspect of the present invention can be achieved.

[0023] The classification apparatus according to each aspect of the present invention may be realized by a computer. In this case, the classification processing program of the classification apparatus that causes the computer to operate as each unit (software element) of the classification apparatus to realize the classification apparatus, and a computer-readable recording medium having the program recorded thereon, are also included in the scope of the present invention. Effect of the Invention

[0024] According to one aspect of the present invention, even beginners can easily distinguish between double folds. [Brief description of the drawings]

[0025] [Figure 1] 1 is an external view of an information processing device and a cell tray according to an embodiment of the present invention; [Diagram 2] 1 is a block diagram showing the functional configuration of an information processing device according to an embodiment of the present invention and a modified example of the embodiment, and the functional configuration of a model generating device according to an embodiment of the present invention. [Diagram 3] 10 is a flowchart showing an example of the flow of discrimination processing by an information processing device according to an embodiment of the present invention. [Figure 4] 1 is an external view of a stock seedling that is the subject of double-petal discrimination by the information processing device. FIG. [Diagram 5] FIG. 2 is an external view showing an example of a display mode of a result of double discrimination in the information processing device. [Figure 6] 5 is a flowchart showing an example of the flow of a model generation process performed by the model generation device. [Figure 7] 1 is a photograph showing a TIF-format image of a target seedling captured in a darkroom in one embodiment of the present invention. [Figure 8] 1 is a photograph showing a JPEG-formatted image of a target seedling captured outdoors in the above embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] Hereinafter, one embodiment of the present invention will be described with reference to Figures 1 to 6. For ease of explanation, members having the same functions as those described in this embodiment will be denoted by the same reference numerals in the modified examples and examples described later, and the description thereof will not be repeated.

[0027] [Outline of the information processing device 100 and the cell tray 500] First, an overview of an information processing device 100 and a cell tray 500 according to an embodiment of the present invention will be described with reference to Fig. 1. The information processing device 100 shown in Fig. 1 is a device capable of processing various information and performing double-leaf discrimination. The information processing device 100 may be, for example, a portable information processing device such as a smartphone or a tablet terminal, or a stationary information processing device such as a desktop PC (Personal Computer). Alternatively, the information processing device 100 may be a seedling sorting robot. In this embodiment, the information processing device 100 is a smartphone.

[0028] There is no limitation on the type of stock that is the target of the double-fold classification by the information processing device 100. Examples of the stock include iron marine, iron white, and iron cherry. Hereinafter, the stock that is the target of the double-fold classification by the information processing device 100 will be referred to as the "target stock."

[0029] As shown in FIG. 1, the information processing device 100 is fixed in position by a fixture 110 so as to be disposed vertically above the cell tray 500. The cell tray 500 is a tray for cultivating a plurality of target stocks by sowing seeds (not shown) of the target stocks. The cell tray 500 has a total of 200 cells 501 in which the seeds of the target stocks are sown, formed in a matrix of 10 rows and 20 columns. In this specification, the up-down direction on the paper surface is the "column" direction, and the left-right direction on the paper surface is the "row" direction. Note that the number and arrangement of the cells 501 in the cell tray 500 are not limited to the above-mentioned embodiment, and cell trays 500 having various numbers and arrangements of the cells 501 formed therein can be used.

[0030] In the cell tray 500, one seed of the target stock is sown in each of the 200 cells 501. When all the seeds grow into seedlings 502 with their cotyledons 503 (see FIG. 4) unfolded, the information processing device 100 is used to identify these seedlings 502 as double-flowered. Hereinafter, the seedlings 502 with their cotyledons 503 unfolded will be referred to as "target seedlings 502." In other words, the target seedlings 502 are unselected seedlings before the information processing device 100 identifies double-flowered seedlings. In the following explanation, the term "cell tray 500" simply refers to the cell tray 500 in a state in which all 200 target seedlings 502 are present.

[0031] For example, a seedling in which the cotyledons 503 have unfolded and the true leaves have begun to appear (about 10 days after sowing in summer) may be used as the target seedling 502, or a seedling in which the true leaves have almost finished appearing (about 16 days after sowing in summer) may be used as the target seedling 502. In general, a seedling in which the cotyledons have unfolded and the true leaves have begun to appear is most suitable for double-petal identification.

[0032] Furthermore, when identifying double-flowered plants, the information processing device 100 is positioned relative to the cell tray 500 so that all 200 target seedlings 502 are displayed on the display screen 101 of the information processing device 100. The display screen 101 of the information processing device 100 specifically refers to the display screen 101 of the display unit 1 (see FIG. 2) provided in the information processing device 100. Hereinafter, of the images displayed on the display screen 101, the image portion showing all 200 target seedlings 502 will be referred to as the "target image."

[0033] The fixture 110 may be a commercially available one or may be one produced for the purpose of double-fold identification by the information processing device 100, and any fixture can be used. In this embodiment, the fixture 110 has a structure that allows it to be attached to the cell tray 500, and the information processing device 100 is fixed to the cell tray 500 by attaching the fixture 110 with the information processing device 100 attached and fixed to the cell tray 500. Note that the fixture 110 may be structured to be placed on the same surface as the surface on which the cell tray 500 is placed, for example.

[0034] The target stock may not be cultivated in the cell tray 500. Seeds of the target stock may be sown in, for example, a pot or directly in a field. The above-mentioned descriptions regarding the information processing device 100 also apply to the information processing device 200 according to the modified example described below.

[0035] [Functional configuration of information processing device 100] Next, a functional configuration of the information processing device 100 will be described with reference to Fig. 2. The information processing device 100 includes a display unit 1, an operation input unit 2, a storage unit 3, an imaging unit 4, and a control unit 5, as shown in Fig. 2.

[0036] The display unit 1 displays the discrimination results by the discrimination device 10 described below and images resulting from the execution of various functions equipped in the information processing device 100 on a display screen 101 (see Figs. 1 and 5). The display unit 1 is an example of a display device according to the present invention. The operation input unit 2 accepts input user operations and outputs the input contents to the control unit 5. Hereinafter, in this specification, the term "user" refers to an operator who performs double discrimination using the information processing device 100. In this embodiment and the modified example described below, the information processing devices 100 and 200 are provided with a touch panel in which the display unit 1 and the operation input unit 2 are integrated. Of course, the display unit 1 and the operation input unit 2 may be separate entities.

[0037] The storage unit 3 stores various programs and data necessary for the information processing device 100 to operate. In addition, a user can store various data and the like in the storage unit 3. The imaging unit 4 is an imaging device that captures images of all 200 target seedlings 502 cultivated in the cell tray 500 to generate captured images. The captured images include both still images and moving images. In this embodiment, the imaging unit 4 is a camera built into a smartphone as the information processing device 100. However, the information processing device 100 does not need to include the imaging unit 4, and for example, a digital camera separate from the information processing device 100 may be used as the imaging device.

[0038] The control unit 5 comprehensively controls each device and each unit provided in the information processing device 100. The control unit 5 has an identification device 10. The identification device 10 is an apparatus for identifying double-flowered seedlings from a total of 200 target seedlings 502, and has a ranking determination unit 11, a judgment unit 12, and a display control unit 13 as shown in Fig. 2. The ranking determination unit 11, the judgment unit 12, and the display control unit 13 will be described in detail later.

[0039] The control unit 5 may be an external device attached to the information processing device 100, or may be a network server used via a communication unit (not shown) of the information processing device 100. Alternatively, the discrimination device 10 may be an external device attached to the information processing device 100.

[0040] [Functional configuration of model generating device 50] Next, the functional configuration of a model generating device 50 according to one embodiment of the present invention will be described with reference to Fig. 2. The model generating device 50 is a device that generates a trained model used in the determination process executed by the determination unit 12. The determination process is a process in which the determination unit 12 inputs predetermined data into the trained model to determine whether or not a specific target seedling 502 is a double-flowered seedling. The details of the determination process will be described later.

[0041] The model generation device 50 may be, for example, a portable information processing device such as a tablet terminal, or a stationary information processing device such as a desktop PC. In addition, for example, the model generation device 50 may be a server such as a cloud. In this embodiment, the model generation device 50 transmits and receives data to and from the information processing device 100 by wireless communication via a communication unit (not shown). As shown in FIG. 2, the model generation device 50 includes a data acquisition unit 51, a learning rank determination unit 52, and a model generation unit 53.

[0042] The data acquisition unit 51 acquires captured images of a plurality of learning seedlings (not shown) and correct answer data corresponding to each of the captured images of the plurality of learning seedlings from the double-leaf discrimination DB 60. The captured images of the learning seedlings are in RAW format and are hereinafter referred to as "learning RAW images." The learning RAW images are still images. The double-leaf discrimination DB 60 is a database in which the learning RAW images and correct answer data are stored.

[0043] The learning seedlings are seedlings in a stock (not shown) after it has been determined whether they are double-flowered seedlings or single-flowered seedlings. Hereinafter, the stock after it has been determined whether the seedlings are double-flowered seedlings or single-flowered seedlings, and whose images are to be stored in the double-flowered identification DB 60, will be referred to as the "learning stock." The correct answer data is data that indicates whether a specific learning seedling is a double-flowered seedling or a single-flowered seedling.

[0044] In this embodiment, the seedlings for learning are mainly stock seedlings that have been subjected to double flower identification mainly by an expert, and stock seedlings that have been cultivated until flowering and have been found to be double flower seedlings or single flower seedlings. In this specification, the term "expert" refers to an operator who can achieve an average correct answer rate of 80% or more in double flower identification. Here, the "correct answer" in "correct answer rate" refers to the fact that the target seedling 502 that an operator (including both beginners and experts) identified as a double flower seedling was actually a double flower seedling. In addition, "average 80% or more" refers to a value obtained by dividing the total correct answer rate of each time by the standard number of times when the operator performs double flower identification a number of times (hereinafter referred to as the standard number of times) that reveals his / her skill level, and the value obtained by dividing the total correct answer rate of each time by the standard number of times is 80% or more.

[0045] However, the learning seedlings are not limited to the above examples. For example, a seedling of a stock that has been subjected to double-petal identification by image analysis using an information processing device other than the information processing device 100 may be used as a learning seedling. As an example of "double-petal identification by image analysis," a method of selecting a seedling of a stock that is relatively light in color by image analysis is envisioned. In addition, there is no particular limitation on the cultivation form of the learning seed stock.

[0046] The learning order determination unit 52 develops the multiple learning RAW images acquired from the data acquisition unit 51 into TIF format. Hereinafter, the learning RAW images developed into TIF format are referred to as "learning TIF images." Then, the learning order determination unit 52 uses the multiple learning TIF images to determine the learning hue angle h-1 and learning saturation C * Then, the training rank determination unit 52 calculates the first training rank of the training hue angle h-1 and the training saturation C * A second learning rank of the learning leaf area s-1 and a third learning rank of the learning leaf area s-1 are determined.

[0047] The learning hue angle h-1 is the hue angle of the cotyledon of the learning seedling. Learning chroma C * -1 is the saturation of the cotyledons of the learning seedlings. The learning leaf area s-1 is the leaf area of ​​the cotyledons of the learning seedlings. Also, the learning hue angle h-1, the learning saturation C * -1 and the training leaf area s-1 are examples of training attribute values ​​according to the present invention. Details of the training TIF images and the method of determining the first to third training ranks will be described later.

[0048] The learning rank determination unit 52 outputs a collection of data sets of the learning TIF images and the first to third learning ranks for each of the multiple learning seedlings as training data to the model generation unit 53. The model generation unit 53 generates a trained model by performing machine learning on the learning model using the training data acquired from the learning rank determination unit 52.

[0049] In this embodiment, the model generating device 50 is configured to transmit and receive data to and from the information processing device 100 by performing wireless communication from outside the information processing device 100. However, the present invention is not limited to this configuration, and the model generating device 50 may be built into the control unit 5 or the classification device 10, for example.

[0050] [Flow of discrimination process by information processing device 100] Next, an example of the flow of the identification process by the information processing device 100 will be described with reference to Figures 3 to 5. The identification process refers to a series of processes from the process in which the imaging unit 4 images all 200 target seedlings 502 and generates captured images, to the process in which the display control unit 13 causes the display unit 1 to highlight a specific image portion on the display screen 101.

[0051] First, the operation input unit 2 accepts a user operation of "start imaging," and the imaging unit 4 images all 200 target seedlings 502 to generate an imaged image in RAW format (hereinafter, "target RAW image"). The imaging unit 4 generates a still image as the target RAW image. At this time, the information processing device 100 is fixed to the cell tray 500 so that the imaging unit 4 can image all 200 target seedlings 502 in the same field of view (see Figures 1 and 5). However, the information processing device 100 may not be fixed to the cell tray 500, and the imaging unit 4 may be able to image each of the 200 target seedlings 502 individually.

[0052] In this embodiment, the cell tray 500 is placed in a darkroom and imaged. However, the cell tray 500 may be placed indoors other than a darkroom and imaged, or may be placed outdoors and imaged. In this specification, a "darkroom" refers to a space in a building where the ambient illuminance is about 500 lux or more and less than about 1,000 lux. Furthermore, an "indoor space other than a darkroom" refers to a space in a building where the ambient illuminance is about 100 lux or more and less than about 2,000 lux. Furthermore, an "outdoor space" refers to a space outside a building where the ambient illuminance is about 2,000 lux or more and less than about 100,000 lux.

[0053] Additionally, "dark room" and "indoors other than a dark room" are examples of locations where artificial light is irradiated to all 200 target seedlings 502 cultivated in the cell tray 500. " " Furthermore, "outdoors" is an example of a location where natural light is irradiated to all 200 target seedlings 502. However, "indoors other than a dark room" also includes the space inside a building as a location where both artificial light and natural light are irradiated to all 200 target seedlings 502.

[0054] After generating the target RAW image, the imaging unit 4 calls up image editing software from the storage unit 3 and develops the target RAW image into TIF format using the software (S11 in FIG. 3). Hereinafter, the target RAW image developed into TIF format will be referred to as the "target TIF image." The target TIF image is an example of target light information of the reflected light from the target seedling 502, and is expressed in the RGB color space. The imaging unit 4 outputs the developed target TIF image to the ranking determination unit 11 of the classification device 10. The imaging unit 4 may develop the target RAW image into, for example, JPEG format or GIF format instead of TIF format.

[0055] Next, the ranking determination unit 11 performs binarization processing on the acquired target TIF image to generate a binarized image. During the binarization processing, the ranking determination unit 11 calculates a hue threshold, a saturation threshold, and a brightness threshold for the boundary between the target seedling 502 and the background in the target TIF image. During the binarization processing, the ranking determination unit 11 also calculates the target leaf area s of the cotyledons 503 for all 200 target seedlings 502 (S12 in FIG. 3).

[0056] The target leaf area s is the surface area of ​​the cotyledon 503, and is an example of a target attribute value according to the present invention. As shown in Fig. 4, since one target seedling 502 has two cotyledons 503, the target leaf area s is the sum of the area of ​​the region surrounded by the outer edge of one cotyledon 503 and the area of ​​the region surrounded by the outer edge of the other cotyledon 503.

[0057] Next, the ranking unit 11 converts the color space of the binarized image from the RGB color space to the CIELAB color space. The CIELAB color space is a nearly complete color space established by the International Commission on Illumination (CIE). It can represent all colors visible to the human eye and can be used as a standard for device-specific models related to image processing.

[0058] The CIELAB color space is * , a * and b * It consists of three coordinates: L * represents the lightness of the color, and L* =0 is black, L * =100 is the diffuse color of white. The reflected color of white is L * = Even higher than 100. * represents the position between red / magenta and green, with negative values ​​being more green and positive values ​​being more red / magenta. * represents a position between yellow and blue, with negative values ​​being bluer and positive values ​​being yellower.

[0059] Next, the ranking determination unit 11 ranks the binarized image in the CIELAB color space by using the threshold values ​​of the three calculated indices (hue, saturation, and brightness). * , a * and b * The three images (hereinafter referred to as "L * Image, a * Image, b * The ranking determination unit 11 may divide the binarized image into a plurality of images (a plurality of images) (S13 in FIG. 3). Note that the ranking determination unit 11 may divide the binarized image using a threshold value of one of the threshold values ​​of the three indices, for example. Furthermore, for example, the ranking determination unit 11 may divide the binarized image using a threshold value different from the threshold values ​​of the three indices.

[0060] Next, the ranking determination unit 11 * Image and b * Using the images, the a of the cotyledon 503 of each of the 200 target seedlings 502 was * value and b * The a value of the cotyledon 503 is calculated (S14 in FIG. 3). * The value is a * Among the pixels constituting the image, a of a plurality of pixels constituting the image area of ​​the cotyledon 503 * The value is the average of the values ​​of cotyledons 503. * The value is b * Among the pixels constituting the image, the b * The values ​​are averaged.

[0061] Next, the ranking determination unit 11 performs the following for each of the 200 target seedlings 502: * value and b *Using the values, the target hue angle h and target saturation C of the cotyledon 503 * (S15 in FIG. 3). The target hue angle h is the hue angle of the cotyledon 503. The ranking determination unit 11 calculates the target hue angle h using the following formula (1). * is the saturation of the cotyledon 503. The ranking determination unit 11 determines the target saturation C * Calculate the target hue angle h and target saturation C * is an example of a target attribute value according to the present invention.

[0062] h=tan -1 (b * / a * ) · · · (1) C * =〔(a * ) 2 +(b * ) 2 〕 1 / 2 (2) The ranking determination unit 11 may convert the color space of the binarized image into, for example, the CIELCh color space instead of the CIELAB color space. In this case, the ranking determination unit 11 converts the C * Images and h * From the image, the target hue angle h and target saturation C of the cotyledon 503 * can be calculated directly.

[0063] Next, the ranking unit 11 ranks the object hue angle h of each of the 200 target seedlings 502 to determine the first object ranking. * The ranking unit 11 determines a second target ranking by ranking the target leaf area s among all 200 target seedlings 502. Furthermore, the ranking unit 11 determines a third target ranking by ranking the target leaf area s among all 200 target seedlings 502 for each of the 200 target seedlings 502 (the above is S16 in FIG. 3: target ranking determination step).

[0064] The ranking determination unit 11 defines the rankings for each of the first to third target rankings as a numerical value, and defines a smaller numerical value as the ranking becomes higher. For example, when the first target ranking of a specific target seedling 502 is the highest, the ranking determination unit 11 defines the first target ranking of the specific target seedling 502 as a numerical value "1". The ranking determination unit 11 also ranks the specific target seedlings by increasing or decreasing the numerical values ​​defining the rankings of the first to third target rankings by "1". For example, when the first target ranking of a specific target seedling 502 is the lowest, the ranking determination unit 11 defines the first target ranking of the specific target seedling 502 as a numerical value "200". In this ranking, it is assumed that the highest first target ranking is defined as a numerical value "1".

[0065] Note that this ranking method is merely an example, and the ranking determination unit 11 can rank the first to third target ranks in various ways. For example, the ranking determination unit 11 may define each of the first to third target ranks so that the higher the rank, the larger the numerical value. Also, for example, the ranking determination unit 11 may rank the first to third target ranks by increasing or decreasing the numerical value defining the rank by "2" (i.e., a natural number other than "1") at a time.

[0066] After completing the ranking, the ranking determination unit 11 outputs the target TIF images and the first to third target rankings of all 200 target seedlings 502 to the judgment unit 12.

[0067] Next, the judgment unit 12 calls the trained model from the model generating device 50 (see Figs. 2 and 6). The judgment unit 12 then inputs the target TIF image and the first to third target rankings of all 200 target seedlings 502 into the trained model, thereby judging whether each of the 200 target seedlings 502 is a double-flowered seedling or not (S17 in Fig. 3: judgment step). The trained model may be stored in advance, for example, in the memory (not shown) of the judgment unit 12 or in the storage unit 3.

[0068] Specifically, the judgment unit 12 acquires, as output results from the trained model, judgment results as to whether each of the 200 target seedlings 502 is a double-flowered seedling or not. The trained model outputs a first judgment result that a specific target seedling 502 is a double-flowered seedling and a second judgment result that a specific target seedling 502 is a single-flowered seedling. Generation of the trained model by the model generation device 50 will be described later.

[0069] The determination unit 12, which has acquired either the first determination result or the second determination result from the trained model for each of the 200 target seedlings 502, outputs these determination results to the display control unit 13. Note that the ranking determination unit 11 may temporarily store the target TIF images and the first to third target rankings in, for example, the storage unit 3. Furthermore, the determination unit 12 may temporarily store the determination results in, for example, the storage unit 3.

[0070] Next, the display control unit 13 causes the display unit 1 to highlight a specific image portion displayed on the display screen 101 based on the determination result acquired from the determination unit 12 (S18 in FIG. 3). In addition to this highlighting, the display control unit 13 controls the display of the display screen 101 by the display unit 1.

[0071] Specifically, as shown in Fig. 5, the display unit 1 displays a ring-shaped mark 102 so as to surround the periphery of the second image portion 104 representing the target seedling 502 that the judgment unit 12 has judged to be a single-flowered seedling (for which the second output result has been output) among the target images displayed on the display screen 101. The display of this mark 102 is performed by the display control of the display control unit 13 based on the judgment result. Such a highlight display allows the user to reliably thin out the seedlings while watching the display screen 101, and the thinning work can be performed simply, quickly and reliably.

[0072] The display control unit 13 can adjust the size and color of the mark 102 on the display unit 1. However, since the user thins out seedlings of a specific target stock while watching the display screen after double-petal discrimination by the information processing device 100 (see Figs. 1 and 5), it is preferable that the size and color of the mark 102 are adjusted within a range that does not interfere with the user's thinning work.

[0073] For example, when reducing the size of the mark 102, it is preferable to reduce the area where the mark 102 overlaps with the second image portion 104 as much as possible so that the seedlings that the user wants to thin out are not difficult to see. Also, for example, when adjusting the color of the mark, it is preferable to avoid excessively fluorescent colors so that the seedlings that the user wants to thin out are not difficult to see. On the other hand, it is also preferable to avoid dark colors so that the user can easily recognize the second image portion 104.

[0074] The highlight display on the display screen 101 by the display control of the display control unit 13 is not limited to the above-mentioned display mode. For example, the display unit 1 may display the mark 102 so as to surround the periphery of the first image portion 103 representing the target seedling 502 that the judgment unit 12 judged to be a double-flowered seedling (output the first output result) among the target images displayed on the display screen 101. Also, for example, the mark 102 may be a hollow triangular or hollow rectangular shape large enough to surround the periphery of the first image portion 103 or the second image portion 104. Furthermore, instead of displaying the mark 102, the first image portion 103 or the second image portion 104 may be illuminated. In this case, it is preferable to illuminate the seedlings that the user wants to thin out so that they are not difficult to see.

[0075] [Flow of model generation process by model generation device 50] Next, an example of the flow of the model generation process by the model generation device 50 will be described with reference to Fig. 6. The model generation process refers to a series of processes from the process in which the data acquisition unit 51 acquires a learning RAW image and correct answer data from the double classification DB 60 to the process in which the model generation unit 53 generates a trained model.

[0076] First, the data acquisition unit 51 acquires a plurality of sets of data sets (hereinafter, "basic data sets") of learning RAW images and correct answer data for one learning seedling from the double-leaf discrimination DB 60. At this time, the data acquisition unit 51 can acquire any number of basic data sets from the double-leaf discrimination DB 60.

[0077] The number of basic data sets to be acquired may be arbitrarily set by, for example, a user operating the model generating device 50 based on the viewpoint of improving the discrimination accuracy of the eight-fold discrimination using the trained model. In this embodiment, the data acquiring unit 51 acquires all basic data sets stored in the eight-fold discrimination DB 60. The data acquiring unit 51 outputs the acquired basic data sets to the learning rank determining unit 52.

[0078] Next, the learning rank determination unit 52 calls up image editing software from a storage unit (not shown) of the model generation device 50, and develops the learning RAW image into a TIF format using the software to generate a learning TIF image (S21: data acquisition step). The learning TIF image is an example of learning light information of the light reflected by the learning seedling, and is expressed in the RGB color space.

[0079] Next, the learning rank determination unit 52 performs binarization processing on the acquired TIF image to generate a binarized image. During the binarization processing, the learning rank determination unit 52 calculates the learning leaf area s-1 of the cotyledons for all of the multiple learning seedlings (S22: data acquisition step). The method of calculating the learning leaf area s-1 is the same as the method of calculating the target leaf area s (see FIG. 4).

[0080] Next, the learning rank determination unit 52 converts the color space of the binarized image from the RGB color space to the CIELAB color space, and converts the binarized image in the CIELAB color space into the L * Image, a * Image, b * Next, the ranking determination unit 11 divides the image into a * Image and b *Using images, the a of cotyledons of multiple test seedlings were * value and b * The a value of the cotyledons of the seedlings for study is calculated (S24). * value and b * The value is the a of the cotyledon 503 of the target seedling 502. * value and b * It is defined similarly to the value.

[0081] Next, the learning order determination unit 52 performs a * value and b * Using the values, the learning hue angle h-1 and learning saturation C of the cotyledons * The learning hue angle h-1 is calculated using the above formula (1) in the same way as the target hue angle h. Also, the learning saturation C * is the target saturation C * Similarly, it is calculated using the above formula (2).

[0082] Next, the learning rank determination unit 52 determines the first learning rank for each of the multiple learning seedlings by ranking the learning hue angle h-1 among the multiple learning seedlings. * -1 among the multiple learning seedlings to determine a second learning rank. Furthermore, the learning rank determination unit 52 determines a third learning rank by ranking the learning leaf area s-1 among the multiple learning seedlings for each of the multiple learning seedlings (S26: learning rank determination step).

[0083] The method of ranking the first to third learning ranks is the same as the method of ranking the first to third target ranks. After completing the ranking, the learning rank determination unit 52 outputs a data set consisting of the learning TIF images of all the multiple learning seedlings and the first to third learning ranks to the model generation unit 53 as teacher data.

[0084] Next, the model generation unit 53 inputs the teacher data acquired from the learning rank determination unit 52 into the learning model, thereby performing machine learning on the learning model to generate a trained model (S27: generation step). The learning model may be stored in, for example, a memory (not shown) of the model generation unit 53, or may be stored in a storage unit of the model generation device 50.

[0085] As the learning model, for example, a CNN (Convolutional Neural Network) can be used. Also, for example, a learning model in which a CNN and a RNN (Recurrent Neural Network) are combined, or a CNN and a LSTM (Long Short-Term Memory) can be combined, can be used.

[0086] A known machine learning algorithm such as multinomial logistic regression analysis may be used to generate the trained model. Alternatively, a known neural network technique (e.g., a known deep learning technique) may be used to generate the trained model. The generated trained model may be temporarily stored in the memory of the model generation unit 53, or in the storage unit of the model generation device 50, for example.

[0087] Next, the model generation unit 53 outputs the trained model to the determination unit 12 of the information processing device 100 by receiving a command to output the trained model from the information processing device 100 (S28). The transmission of the command from the information processing device 100 is triggered, for example, by the operation input unit 2 of the information processing device 100 receiving a user operation to "start imaging."

[0088] The model generating device 50 appropriately executes each process of S21 to S27 in the flowchart of FIG. 6, triggered by the storage of a new basic data set in the double discrimination DB 60, a change in the data configuration of the teacher data, or the like.

[0089] [Modifications] <Modification of the information processing device 100> (Number and type of target rankings) The configuration and processing of the information processing device 100 described above are merely examples, and multiple different configurations and processing can be adopted. First, for example, the judgment unit 12 of the information processing device 100 may input any one or any two of the first to third target orders into the trained model to determine whether a specific target seedling 502 is a double-flowered seedling or not. By reducing the number of input data to the trained model in this way, the pre-processing of double-flowered discrimination can be simplified. In this case, the first to third learning orders included in the teacher data are also set to any one or any two learning orders corresponding to the input data to the trained model.

[0090] Also, for example, the determination unit 12 of the information processing device 100 may input one or more other object rankings to the trained model in addition to the first to third object rankings. Increasing the number of input data to the trained model in this way can improve the determination accuracy of the determination unit 12. As the other object ranking, for example, an object ranking determined by ranking the color difference of the cotyledons 503 can be assumed. This other object ranking may be used in place of all of the first to third object rankings, or may be used by replacing any one of the first to third object rankings. The increase or decrease in the number of the first to third learning rankings included in the teacher data is as described above.

[0091] (Using a color sensor) As a modified example of the information processing device 100, for example, an information processing device 200 shown in FIG. 2 can be assumed. The information processing device 200 includes a color sensor 6 instead of the imaging unit 4. That is, in the information processing device 200, the target light information of the reflected light from the target seedling 502 is not a TIF image but detection information detected by the color sensor 6. By using the color sensor 6, the target light information is a * value and b * The process of obtaining the value is simplified, and the processing cost can be reduced.

[0092] The color sensor 6 has a light source and a light receiving unit (both not shown), and is configured to detect the reflected light that is emitted from the light source and reflected by the target seedling 502 with the light receiving unit. The color sensor 6 may output RGB values, * value and b * In the case where the color sensor 600 outputs RGB values, the ranking determination unit 11 determines a * value and b * Calculate the value.

[0093] a * value and b * A specific example of the color sensor 6 that directly outputs a value is a color difference meter (CR-200 manufactured by Konica Minolta). * value and b * The method for obtaining the value is as follows. The light source of the color difference meter is C or D. 65 Before starting the measurement, white calibration is performed using a white calibration plate. Measurements of all the target seedlings 502 are performed under the same setting and calibration conditions. The target seedlings 502 are placed on the white calibration plate, and a color difference meter that has been set and calibrated is used to measure the color difference between the target seedlings 502 and the white calibration plate. * value and b * Measure the value.

[0094] (How to display the judgment results) Furthermore, the information processing device 100 does not require the display control unit 13 to highlight the judgment result of the judgment unit 12 on the display screen 101. For example, after the judgment process of the judgment unit 12 is completed, the operation input unit 2 may accept a user operation to "display the judgment result", and the display control unit 13 may cause the display unit 1 to display the judgment result in text (for example, "Seedling No. 10: double-flowered seedling") on the display screen 101. Note that this text display may be displayed on the display screen of an information processing device other than the information processing device 100. In this case, the display control unit 13 instructs the display unit of the information processing device other than the information processing device 100 to display the text.

[0095] <Modifications of the model generating device 50> The configuration and processing of the model generating device 50 described above are merely examples, and multiple different configurations and processing can be adopted. For example, the model generating device 50 may have the learning rank determining unit 52 develop a first learning TIF image and a second learning TIF image as learning TIF images constituting the teacher data. The first learning TIF image is an example of first learning light information of a first reflected light in which artificial light irradiated in a darkroom is reflected by the learning seedlings. The second learning TIF image is an example of second learning light information of a second reflected light in which natural light irradiated outdoors is reflected by the learning seedlings.

[0096] By including these two types of training TIF images in the teacher data, the trained model can correctly determine whether the target seedling 502 to be subjected to double-flowered discrimination is a double-flowered seedling or not with a certain level of accuracy, regardless of whether the target seedling 502 is placed in a darkroom or outdoors. Note that when the training rank determination unit 52 develops the first training TIF image and the second training TIF image, it is assumed that the double-flowered discrimination DB 60 stores a training RAW image captured in a darkroom and a training RAW image captured outdoors.

[0097] [Software implementation example] The control block of the discrimination device 10 (particularly the ranking unit 11 and the judgment unit 12) may be realized by a logic circuit (hardware) formed on an integrated circuit (IC chip) or the like, or may be realized by software.

[0098] In the latter case, the classification device 10 is equipped with a computer that executes instructions of a program (classification processing program), which is software that realizes each function. This computer is equipped with, for example, one or more processors, and a computer-readable recording medium that stores the program. Then, in the computer, the processor reads the program from the recording medium and executes it, thereby achieving the object of the present invention.

[0099] The processor may be, for example, a CPU (Central Processing Unit). The recording medium may be a "non-transient tangible medium", such as a ROM (Read Only Memory), tape, disk, card, semiconductor memory, programmable logic circuit, or the like. The computer may further include a RAM (Random Access Memory) for expanding the program. The program may be supplied to the computer via any transmission medium (such as a communication network or broadcast waves) capable of transmitting the program. One aspect of the present invention may also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission.

[0100] [Example] Next, Examples 1 to 3 of the present invention will be described with reference to Figures 7 and 8. In Examples 1 to 3 and Comparative Examples 1 to 3, as shown in Figures 7 and 8, a seed of the target stock (iron marine) was sown one by one in all 200 cells 501 formed in a cell tray 500, and the seedlings were raised under mist watering to cultivate the target seedlings 502. Furthermore, a PC (model number: NG-im620GA1, manufactured by Mouse Computer Co., Ltd.) was used as the information processing device 100, and an "iPhone 6s (registered trademark; manufactured by Apple Inc.)" was used as the imaging device instead of the imaging unit 4.

[0101] <Image capture method> In Examples 1 to 3 and Comparative Examples 1 to 3, when the cotyledons 503 unfolded and the true leaves began to appear, all 200 target seedlings 502 were photographed from the vertically upper side of the cell tray 500 to generate photographed images of the target seedlings 502 and photographed images of the learning seedlings. In addition, images were taken of the same cell tray 500 in two locations, a dark room and outdoors. The ambient illuminance in the dark room was about 500 lux. The ambient illuminance outdoors was about 20,000 to about 100,000 lux.

[0102] The image of the target seedling 502 captured in a darkroom was in RAW format. Similarly, the image of the learning seedling captured in a darkroom was also in RAW format (hereinafter, "learning RAW image"). The image of the target seedling 502 captured outdoors was in JPEG format, not RAW format. Similarly, the image of the learning seedling captured outdoors was also in JPEG format, not RAW format (hereinafter, "learning JPEG image").

[0103] In the darkroom, the camera of the information processing device 100 was not used, but instead a digital camera (Nikon D3200; manufactured by Nikon Corporation) was used to capture the image. When capturing the image, the digital camera was fixed in position relative to the cell tray 500 so that the lens surface of the digital camera was approximately parallel to the upper surface of the cell tray 500 (the surface on which the cells 501 were formed). A fixture (not shown) dedicated to digital cameras was used to fix the position. In addition, a reflector lamp (PRF-500WB / D, color temperature 500K; manufactured by Panasonic Corporation) was used as a light source to illuminate the cell tray 500. The photograph shown in FIG. 7 is an example of an image obtained by developing a RAW-format image of the target seedling 502 captured in a darkroom into a TIF format.

[0104] On the other hand, outdoors, the camera of the information processing device 100 was not used, but the camera of the iPhone 6s (registered trademark; manufactured by Apple Inc.) was used to capture the image. The iPhone 6s (registered trademark; manufactured by Apple Inc.) was not fixed in position by a fixture 110 or the like, and the user held the iPhone 6s (registered trademark; manufactured by Apple Inc.) to capture the image. When capturing the image, the lens surface of the camera and the upper surface of the cell tray 500 were made approximately parallel. In addition, the image was captured under natural light without using an artificial light source such as the above-mentioned reflector lamp. The photograph shown in FIG. 8 is an example of a JPEG-format captured image of the target seedling 502 captured outdoors.

[0105] <How to generate a trained model> In Examples 1 to 3, a trained model was generated using training data consisting of a plurality of data sets, each set consisting of a training JPEG image, first to third training ranks determined based on the training JPEG image, and correct answer data. For the training JPEG images in Examples 1 to 3, images were used that were obtained by developing training RAW images of iron marine seedlings into TIF format and then converting them into JPEG format, and JPEG images.

[0106] A CNN was used as a learning model, and the training data was input to the CNN to perform machine learning on the CNN. For the machine learning, "TensorFlow (registered trademark; manufactured by Google Inc.)", an open software library, was used. In addition, the machine learning and generation of the trained model were performed by installing TensorFlow on a PC (model number: NG-im620GA1, manufactured by Mouse Computer Co., Ltd.), which is also the model generation device 50.

[0107] On the other hand, in Comparative Example 3, one set of data includes a learning JPEG image and a learning hue angle h-1 and learning saturation C * A trained model was generated using training data consisting of the absolute values ​​of the leaf area s-1 and the learning leaf area s-1 and the correct answer data. The rest is the same as in Examples 1 to 3.

[0108] <Double-fold identification and results> (Example of double-fold discrimination) First, as Example 1, teacher data was generated using 23 learning JPEG images captured in a darkroom. Next, in the model generating device 50, a trained model was generated by machine learning the CNN using the teacher data. Next, in the information processing device 100, 14 target JPEG images of the target seedling 502 captured in a darkroom and the corresponding first to third target ranks were input to the trained model, and a judgment result (discrimination result) was obtained for each of the 14 target JPEG images. As described above, a PC (model number: NG-im620GA1, manufactured by Mouse Computer Co., Ltd.) was used as the model generating device 50 and also as the information processing device 100.

[0109] In addition, as Example 2, training data was generated using 32 training JPEG images captured in a darkroom and outdoors. Of the total 32 training JPEG images, 20 training JPEG images were captured in a darkroom and 12 training JPEG images were captured outdoors.

[0110] Next, in the model generating device 50, a trained model was generated by machine learning CNN using this teacher data. Next, in the information processing device 100, 23 target JPEG images of the target seedlings 502 captured in a darkroom and outdoors, and the corresponding first to third target rankings were input to the trained model, and a judgment result (discrimination result) was obtained for each of the 23 target JPEG images. Of the total 23 target JPEG images, 5 target JPEG images were captured in a darkroom and 18 target JPEG images were captured outdoors.

[0111] Furthermore, as Example 3, training data was generated using 142 training JPEG images captured in a darkroom and outdoors. Of the 142 training JPEG images, 65 training JPEG images were captured in a darkroom and 77 training JPEG images were captured outdoors.

[0112] Next, in the model generating device 50, a trained model was generated by machine learning CNN using this teacher data. Next, in the information processing device 100, 54 target JPEG images of the target seedlings 502 captured in a darkroom and outdoors, and the corresponding first to third target rankings were input to the trained model, and a judgment result (discrimination result) was obtained for each of the 54 target JPEG images. Of the total 54 target JPEG images, 21 target JPEG images were captured in a darkroom, and 33 target JPEG images were captured outdoors.

[0113] (Comparative Example of Double Identification) First, as Comparative Example 1, for each of two target TIF images captured in a darkroom, the saturation of the cotyledons of all 200 target seedlings 502 was calculated using the target TIF images. The above-mentioned formula (2) was used to calculate this saturation. Next, for each calculated saturation, the target seedlings 502 were ranked among all 200 target seedlings, and the target seedlings 502 included in the top 50% were determined (classified) as double-flowered seedlings, and the target seedlings 502 included in the bottom 50% were determined (classified) as single-flowered seedlings.

[0114] In Comparative Example 2, the saturation of the cotyledons of all 200 target seedlings 502 was calculated for each of two target JPEG images captured outdoors using the target JPEG images. The saturation was calculated using the above formula (2) as in Comparative Example 1. Next, the calculated saturations were ranked among all 200 target seedlings 502, and the target seedlings 502 in the top 50% were determined (classified) as double-flowered seedlings, and the target seedlings 502 in the bottom 50% were determined (classified) as single-flowered seedlings. In Comparative Examples 1 and 2, the above-mentioned series of processes were performed using a PC different from the PC (model number: NG-im620GA1, manufactured by Mouse Computer Co., Ltd.) that is the information processing device 100.

[0115] Furthermore, as a comparative example 3, training data was generated using 12 JPEG images for training captured in a darkroom. Next, in the model generation device 50, a trained model was generated by machine learning the CNN using this training data. Next, in the information processing device 100, five target JPEG images of the target seedling 502 captured in a darkroom and the corresponding target hue angle h and target saturation C were generated. * By inputting the absolute values ​​of the target leaf area s and the target leaf area s into the trained model, judgment results (identification results) were obtained for each of the five JPEG images.

[0116] (Identification results) As for Examples 1 to 3, as shown in Table 1 below, the final correct answer rate of the classification result of Example 1 was 83%. Moreover, the final correct answer rate of the classification result of Example 2 was 76%. Furthermore, the final correct answer rate of the classification result of Example 3 was 84%. As for Comparative Examples 1 to 3, as shown in Table 2 below, the final correct answer rate of the classification result of Comparative Example 1 was 71%. Moreover, the final correct answer rate of the classification result of Comparative Example 2 was 68%. Furthermore, the final correct answer rate of the classification result of Comparative Example 3 was 74%.

[0117] The final correct answer rate for each Example and Comparative Example was calculated using the following formulas (3) to (5). Note that the "number of learning images" in Tables 1 and 2 corresponds to the number of learning JPEG images. Also, the "number of test images" in the same tables corresponds to the number of target TIF images or JPEG images. Furthermore, the "correct answer rate" in the same tables corresponds to the final correct answer rate.

[0118] [Table 1]

[0119] [Table 2]

[0120] Correct answer rate for double-flowered seedlings = (number of target seedlings 502 that were determined to be double-flowered seedlings and were actually double-flowered seedlings) / ((number of target seedlings 502 that were determined to be double-flowered seedlings) - (number of target seedlings 502 that were determined to be double-flowered seedlings and died)) × 100 [%] · · · (3) Correct answer rate for single-flowered seedlings = (number of target seedlings 502 that were determined to be single-flowered seedlings and were actually single-flowered seedlings) / ((number of target seedlings 502 that were determined to be single-flowered seedlings) - (number of target seedlings 502 that were determined to be single-flowered seedlings and died)) × 100 [%] (4) Final correct answer rate = ((correct answer rate for double-flowered seedlings) - (correct answer rate for single-flowered seedlings)) / 2[%] (5) (Consideration) In general, the accuracy of double-petal identification using an image of a stock seedling taken in a darkroom is higher than that of double-petal identification using an image of a stock seedling taken outdoors. This is because the color of the seedling reflected in an image taken outdoors changes significantly depending on the weather, shooting location, etc., even if the subject is the same. However, as shown in the above identification results, even Example 2 (final correct answer rate: 76%), which had the lowest final correct answer rate and included 18 target JPEG images taken outdoors, had a higher final correct answer rate than Comparative Example 3 (final correct answer rate: 74%), which had the highest final correct answer rate.

[0121] From this, it was found that the information processing device according to one embodiment of the present invention can perform double-fold classification with the same or higher classification accuracy as the conventional double-fold classification using a target JPEG image captured in a darkroom as the classification target, even if the target JPEG image captured outdoors is included in the classification target. In addition, it was found that the classification accuracy is improved by ranking each absolute value within a predetermined group and converting it into a relative value, rather than using the absolute values ​​of the hue angle, saturation, and leaf area as parameters constituting the teacher data and input data to the trained model.

[0122] Furthermore, Example 3, in which the conditions for generating the trained model (learning conditions) were substantially the same as those in Example 2 except that the number of training JPEG images was increased from Example 2 (32 images) (142 images), had a higher final correct answer rate than Example 2 (final correct answer rate: 84%). Moreover, the final correct answer rate in Example 3 was substantially the same as that in Example 1 (final correct answer rate: 83%), in which only training JPEG images and target JPEG images captured in a darkroom were used. From these findings, it was found that by increasing the number of training JPEG images, the classification accuracy of the information processing device according to one embodiment of the present invention can be made equal to or higher than that of conventional double classification, even when training JPEG images and target JPEG images captured outdoors are included.

[0123] [Additional Notes] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in the embodiments and modifications are also included in the technical scope of the present invention. [Explanation of symbols]

[0124] 1 Display section (display device) 10 Identification device 11 Ranking Division 12 Judgment section 13 Display control section 100, 200 Information processing device 103 First image part 104 Second image part 502 Target seedlings (target stock seedlings) 503 Cotyledons

Claims

1. 1. A discrimination device for discriminating a double-flowered seedling from a plurality of target stock seedlings before discriminating the double-flowered seedling, A trained model is generated by machine learning using training data including: (i) training light information of reflected light reflected by the seedlings; (ii) training ranking determined by ranking the training attribute values ​​of the cotyledons among the seedlings in the training stocks after it has been determined whether the seedlings are double-flowered seedlings; and (iii) correct answer data indicating whether the seedlings are double-flowered seedlings. a ranking determination unit that acquires a target attribute value of a cotyledon from target light information of light reflected by each of the seedlings of the plurality of target stocks, and ranks the target attribute values ​​among the seedlings of the plurality of target stocks to determine a target ranking; A determination unit that determines whether the seedling of the specific target stock is the double-flowered seedling by inputting the target light information and the target ranking of the seedling of the specific target stock into the trained model, The learning light information included in the teacher data includes the following (1) or (2): (1) first learning light information of a first reflected light of artificial light reflected by a seedling of the learning stock; (2) The first learning light information and second learning light information of second reflected light, which is natural light reflected by the seedlings of the learning stock.

2. The learning rankings included in the teacher data include: (i) a first learning ranking determined by ranking the learning hue angles of the cotyledons among the seedlings of the plurality of learning stocks; (ii) a second learning ranking determined by ranking the learning saturation of the cotyledons among the seedlings of the plurality of learning stocks; and (iii) a third learning ranking determined by ranking the learning leaf area of ​​the cotyledons among the seedlings of the plurality of learning stocks. The ranking determination unit, for each of the plurality of target stock seedlings, From the target light information, a target hue angle, a target saturation, and a target leaf area of ​​the cotyledons are obtained as the target attribute values; (α) ranking the object hue angle among the plurality of object stock seedlings to determine a first object ranking, and (β) ranking the object saturation among the plurality of object stock seedlings to determine a second object ranking, and (γ) ranking the object leaf area among the plurality of object stock seedlings to determine a third object ranking; The identification device described in claim 1, wherein the judgment unit determines whether the specific target stock seedling is a double-flowered seedling by inputting the target light information, the first target ranking, the second target ranking, and the third target ranking of the specific target stock seedling into the learned model.

3. The discrimination device according to claim 1 , wherein the learning light information included in the teacher data includes the first learning light information and the second learning light information.

4. A display control unit controls a display device to display target images representing the seedlings of the plurality of target stocks on a display screen, An appraisal device as described in any one of claims 1 to 3, wherein the display control unit highlights either a first image portion representing a seedling of the target stock that has been determined to be a double-flowered seedling by the judgment unit, or a second image portion representing a seedling of the target stock that has been determined to be a single-flowered seedling by the judgment unit, of the target image displayed on the display screen.

5. The first reflected light is the artificial light irradiated in a dark room or a room other than the dark room and reflected by the seedlings of the stock for study, 4. The identification device according to claim 3, wherein the second reflected light is natural light irradiated outdoors and reflected by the seedlings of the training stock.

6. An information processing device comprising the discrimination device according to claim 1 .

7. 2. A discrimination processing program for causing a computer to function as the discrimination device according to claim 1, the discrimination processing program causing a computer to function as the ranking determination unit and the judgment unit.

8. A method for generating training data for generating a trained model, which is applied to a classification device that performs machine learning to classify double-flowered seedlings from the seedlings of the plurality of target stocks, the method comprising: a ranking determination unit that acquires target attribute values ​​of cotyledons from target light information of reflected light reflected by the seedlings for each of the seedlings of a plurality of target stocks before the double-flowered seedlings are classified; and a ranking determination unit that determines a target ranking by ranking the target attribute values ​​among the seedlings of the plurality of target stocks; The method for generating teacher data is executed by a computer, The method for generating teacher data includes: In the teacher data, in a plurality of learning stock seedlings after it has been determined whether the seedlings are double-flowered or not, Learning light information of the reflected light from the seedling; A learning rank determined by ranking the learning attribute values ​​of the cotyledons among the seedlings of the plurality of learning stocks; and correct answer data indicating whether the seedling is a double-flowered seedling or not. The teacher data is used in a process in which the discrimination device determines whether the seedling of the specific target stock is the double-flowered seedling by inputting the target light information and the target ranking of the seedling of the specific target stock into the trained model, The method for generating teacher data includes: A method for generating teacher data, comprising the step of including the following (1) or (2) in the learning light information included in the teacher data: (1) first learning light information of a first reflected light of artificial light reflected by a seedling of the learning stock; (2) The first learning light information and second learning light information of second reflected light, which is natural light reflected by the seedlings of the learning stock.

9. A method for identifying a double-flowered seedling from a plurality of target stock seedlings before identifying the double-flowered seedling, comprising: A trained model is generated by machine learning using training data including: (i) training light information of reflected light reflected by the seedlings; (ii) training ranking determined by ranking the training attribute values ​​of the cotyledons among the seedlings in the training stocks after it has been determined whether the seedlings are double-flowered seedlings; and (iii) correct answer data indicating whether the seedlings are double-flowered seedlings. a target ranking determination step of acquiring a target attribute value of a cotyledon from target light information of light reflected by each of the seedlings of the plurality of target stocks, and ranking the target attribute values ​​among the seedlings of the plurality of target stocks to determine a target ranking; and a determination step of determining whether the seedling of the specific target stock is the double-flowered seedling by inputting the target light information and the target ranking of the seedling of the specific target stock into the trained model, The learning light information included in the teacher data includes the following (1) or (2): (1) first learning light information of a first reflected light of artificial light reflected by a seedling of the learning stock; (2) The first learning light information and second learning light information of second reflected light, which is natural light reflected by the seedlings of the learning stock.

10. A method for generating a trained model for use in a discrimination device for discriminating double-flowered seedlings from seedlings of a plurality of target stocks before discriminating the double-flowered seedlings, comprising: a data acquisition step of acquiring (i) learning light information of light reflected by the seedlings, (ii) learning attribute values ​​of the cotyledons, and (iii) correct answer data indicating whether the seedlings are double-flowered seedlings or not, for the seedlings of the multiple learning stocks after it has been determined whether they are double-flowered seedlings or not; a learning rank determination step of determining a learning rank for each of the plurality of learning stock seedlings by ranking the learning attribute values ​​among the plurality of learning stock seedlings; A generation step of inputting a data set including the learning light information, the correct answer data, and the learning ranking into a learning model as teacher data, and generating the trained model by machine learning the learning model, A generation method of the learning light information included in the teacher data, the learning light information including the following (1) or (2): (1) first learning light information of a first reflected light of artificial light reflected by a seedling of the learning stock; (2) The first learning light information and second learning light information of second reflected light, which is natural light reflected by the seedlings of the learning stock.

Images