EXCRETION PROPERTY ESTIMATION MODEL LEARNING DEVICE, EXCRETION PROPERTY ESTIMATION DEVICE, EXCRETION PROPERTY ESTIMATION MODEL LEARNING METHOD, EXCRETION PROPERTY ESTIMATION METHOD AND PROGRAM

Abstract

To estimate the property of excrement of an animal with higher accuracy.SOLUTION: An excrement property estimation model learning device comprises: a learning data acquisition unit which acquires learning data including input side learning data including an excrement space being a predetermined space and being a space where excrement of a target animal of breeding exists and image data for learning being image data of a photographing result image obtained by imaging the excrement space and a correct answer label indicating whether or not abnormal excrement is reflected in the image indicated by the image data for learning; an excrement property estimation learning model execution unit which estimates such a possibility that the abnormal excrement is reflected in the image indicated by the image data for learning by using the excrement property estimation learning model being the machine learning model that estimates such a possibility that the abnormal excrement is reflected in the image data for learning; and an update unit which updates the excrement property estimation learning model so as to reduce the difference on the basis of the difference between the estimation result of the excrement property estimation learning model execution unit and the correct answer label.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an excrement property estimation model learning device, an excrement property estimation device, an excrement property estimation model learning method, an excrement property estimation method, and a program. [Background technology]

[0002] When raising experimental animals such as marmosets, the characteristics of their feces can be used to understand the animal's health condition and psychological stress. In such cases, the breeder determines the characteristics of the feces based on the color and shape of the feces, and also checks for the presence or absence of vomit. However, the breeder's ability to determine the characteristics of the feces is greatly influenced by the breeder's skill, and the results of the determination may vary depending on the breeder's level of proficiency. Therefore, attempts are sometimes made to determine the characteristics of the feces using computer programs for image analysis. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2008-48683 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, in animal cages, not only feces and vomit but also food is often found. Moreover, the color and shape of food are often similar to those of feces. Therefore, in conventional image analysis programs that attempt to determine the properties of feces only by comparing with a standard previously set by a programmer or user, there are cases where it is not possible to distinguish between food and feces. For example, conventional image analysis programs estimate the area of ​​feces and determine that the feces is abnormal when an area equal to or exceeds a threshold value. As a result, when an animal is in an abnormal state, such as in a poor health state or under strong psychological stress, feces and food are not distinguished from each other, and the properties of the feces may not be properly estimated. In addition, not only when feces and food are distinguished from each other when an animal is in an abnormal state, but also when the animal is in a normal state, feces and food are not distinguished from each other, and the properties of the feces may not be properly estimated. In addition, this is a common problem not only for feces but also for excrement such as vomited matter.

[0005] In view of the above circumstances, an object of the present invention is to provide a technique for estimating the properties of animal excrement with higher accuracy. [Means for solving the problem]

[0006] One aspect of the present invention is an excrement property estimation model learning device that includes: an excrement space, which is a predetermined space in which excrement of a kept animal is present; a learning data acquisition unit that acquires learning data including input side learning data including learning image data, which is image data of a photographed image of the excrement space, and a correct answer label indicating whether or not an image represented by the learning image data contains abnormal excrement, an excrement property estimation learning model execution unit that estimates the probability that an image represented by the learning image data contains abnormal excrement using an excrement property estimation learning model, which is a machine learning model that estimates the probability that abnormal excrement is depicted in the learning image data, and an update unit that updates the excrement property estimation learning model based on a difference between the estimation result of the excrement property estimation learning model execution unit and the correct answer label so as to reduce the difference.

[0007] One aspect of the present invention is the above-mentioned excrement property estimation model learning device, wherein the learning image data is image data of an extracted excrement image, which is an image extracted in advance from a shooting result image and is an image of excrement having a horizontal projection area equal to or greater than a predetermined area.

[0008] One aspect of the present invention is the above-mentioned excrement property estimation model learning device, wherein the learning image data is image data of an image after image processing has been performed on the captured image using a filter that does not pass the colors of obstacles other than excrement whose color and shape are within a predetermined reference value.

[0009] One aspect of the present invention is the excrement property estimation model learning device, wherein the input side learning data includes information indicating a date and time when the learning image data was captured.

[0010] One aspect of the present invention is the excrement property estimation model learning device, wherein the input side learning data includes information indicating the maximum number of animals from which the excrement present in the excrement space belongs.

[0011] One aspect of the present invention is the above-mentioned excrement property estimation model learning device, wherein the input side learning data includes information indicating the feed given to the animal to be kept at a predetermined timing before the time when an image is captured by an imaging device to obtain a captured image.

[0012] According to one aspect of the present invention, there is provided an excrement space that is a predetermined space where excrement from a bred animal exists, an analysis target data acquisition unit that acquires analysis target data including image data of a photographed image of the excrement space, a learning data acquisition unit that acquires learning data including input side learning data including learning image data that is image data of the photographed image, and a correct answer label that indicates whether or not an abnormal excrement is captured in an image shown by the learning image data, and an excrement property estimation unit that estimates the probability that an abnormal excrement is captured in the learning image data. The excrement property estimation device includes an estimation unit that estimates the probability that abnormal excrement is included in the image data acquired by the analysis target data acquisition unit using a trained excrement property estimation learning model obtained by an excrement property estimation model learning device that includes an excrement property estimation learning model execution unit that uses an estimation learning model to estimate the probability that abnormal excrement is included in the image represented by the learning image data, and an update unit that updates the excrement property estimation learning model based on the difference between the estimation result of the excrement property estimation learning model execution unit and the correct label so as to reduce the difference.

[0013] One aspect of the present invention is an excrement property estimation model learning method comprising: a learning data acquisition step of acquiring learning data including an excrement space, which is a predetermined space in which excrement of a kept animal is present; input side learning data including learning image data, which is image data of a resultant image obtained by photographing the excrement space; and a correct answer label indicating whether or not an abnormal excrement is depicted in an image represented by the learning image data; an excrement property estimation learning model execution step of estimating the probability that an abnormal excrement is depicted in an image represented by the learning image data, using an excrement property estimation learning model, which is a machine learning model that estimates the probability that abnormal excrement is depicted in the learning image data; and an update step of updating the excrement property estimation learning model based on a difference between the estimation result of the excrement property estimation learning model execution step and the correct answer label so as to reduce the difference.

[0014] According to one aspect of the present invention, there is provided an excrement space, which is a predetermined space in which excrement from a bred animal exists, and an analysis target data acquisition step for acquiring analysis target data including image data of a photographed image of the excrement space; a learning data acquisition unit for acquiring learning data including input side learning data including learning image data which is image data of the photographed image, and a correct answer label indicating whether or not an abnormal excrement is captured in an image represented by the learning image data; and an excrement property estimation unit for estimating the probability that an abnormal excrement is captured in the learning image data. the excrement property estimation model executing unit estimating the probability that abnormal excrement is captured in the image data obtained in the analysis target data acquisition step using a trained excrement property estimation learning model obtained by an excrement property estimation model learning device having: an excrement property estimation learning model executing unit estimating the probability that abnormal excrement is captured in the image represented by the learning image data using a fixed learning model; and an updating unit updating the excrement property estimation learning model based on the difference between the estimation result of the excrement property estimation learning model executing unit and the correct label so as to reduce the difference.

[0015] One aspect of the present invention is a program for causing a computer to function as the excrement property estimation model learning device.

[0016] One aspect of the present invention is a program for causing a computer to function as the excrement property estimation device. Effect of the Invention

[0017] According to the present invention, it is possible to estimate the properties of animal excrement with higher accuracy. [Brief description of the drawings]

[0018] [Figure 1] FIG. 1 is an explanatory diagram illustrating an outline of an excrement property estimation system 100 according to an embodiment. [Diagram 2] FIG. 4 is a diagram showing a first example of a photographed image of a feces space in an embodiment. [Diagram 3]FIG. 11 is a diagram showing a second example of a photographed image of a feces space in the embodiment. [Figure 4] FIG. 11 is a diagram showing a third example of a photographed image of a feces space in the embodiment. [Diagram 5] FIG. 11 is a view showing a fourth example of a photographed image of a feces space in the embodiment. [Figure 6] FIG. 11 is a diagram showing a fifth example of a photographed image of a feces space in an embodiment. [Figure 7] FIG. 13 is a diagram showing a sixth example of a photographed image of a feces space in an embodiment. [Figure 8] FIG. 13 is a diagram showing a seventh example of a photographed image of a feces space in an embodiment. [Figure 9] FIG. 13 is a diagram showing an eighth example of a photographed image of a feces space in an embodiment. [Figure 10] FIG. 13 is a diagram showing a ninth example of a photographed image of a feces space in an embodiment. [Figure 11] FIG. 2 is a diagram showing an example of the hardware configuration of the fecal property estimation model learning device 1 according to the embodiment. [Figure 12] FIG. 2 is a diagram showing an example of the functional configuration of a control unit 11 in the embodiment. [Figure 13] 4 is a flowchart showing an example of the flow of processing executed by the fecal characteristics estimation model learning device 1 of the embodiment. [Figure 14] FIG. 2 is a diagram showing an example of the hardware configuration of a feces property estimation device 2 in an embodiment. [Figure 15] FIG. 2 is a diagram showing an example of the functional configuration of a control unit 21 in the embodiment. [Figure 16] 4 is a flowchart showing an example of the flow of processing executed by a fecal property estimating device 2 in an embodiment. [Figure 17] 11 is a diagram showing an example of an experiment result of estimating abnormal stool using the excrement property estimation system 100 of the embodiment. FIG. [Figure 18] FIG. 13 is a diagram showing an example of the hardware configuration of a feces property estimation device 2a in a modified example. [Figure 19] FIG. 13 is a diagram showing an example of a functional configuration of a control unit 21a in a modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] (Embodiment) FIG. 1 is an explanatory diagram for explaining an overview of an excrement property estimation system 100 according to an embodiment. The term "excrement" used in this specification encompasses feces and vomit. For the sake of simplicity, the excrement property estimation system 100 will be described below using an example in which the excrement is feces from a domestic animal 900, but the excrement is not limited to feces. That is, the processing performed on feces in the following description may also be performed on other excrement such as vomit. The excrement may also be vomit from a domestic animal 900.

[0020] The excrement property estimation system 100 obtains information (hereinafter referred to as "relationship information") indicating a relationship between an image (hereinafter referred to as "photographed image") of the feces space and the probability that abnormal feces is captured in the image by a machine learning method. The feces space is a predetermined space in which at least feces of the animal being kept exists when the animal being kept excretes. The space between the floor mesh 902 and the bedding 903 in FIG. 1 is an example of the feces space. Abnormal feces is feces that is abnormal. Hereinafter, the animal being kept is referred to as a kept animal 900. The kept animal 900 is, for example, a laboratory animal, such as a marmoset. The kept animal 900 may be a laboratory animal such as a mouse, a rat, a guinea pig, a dog, a rabbit, or a rhesus monkey.

[0021] The excrement property estimation system 100 obtains relevant information by updating the feces property estimation learning model by a machine learning method. The feces property estimation learning model is a machine learning model that estimates the probability that abnormal stool is captured in a photographed image based on the result of photographing the feces space.

[0022] A machine learning model (hereinafter referred to as a "machine learning model") is a mathematical model having one or more processes whose execution conditions and order (hereinafter referred to as "execution rules") are predetermined. For simplicity of the following explanation, updating a machine learning model using a machine learning method is referred to as learning. Furthermore, updating a machine learning model means suitably adjusting parameter values ​​in the machine learning model. Furthermore, executing a machine learning model means executing each process included in the machine learning model in accordance with the execution rules.

[0023] The machine learning model is expressed, for example, by a neural network. A neural network is a circuit such as an electronic circuit, an electric circuit, an optical circuit, or an integrated circuit that expresses a machine learning model. The parameters of the neural network are suitably adjusted based on the value of the objective function (i.e., the loss). The parameters of the network are the parameters of the circuits that constitute the network. Such a neural network model that is multi-layered is called a deep learning model. A method of learning a machine learning model is, for example, deep learning.

[0024] The farmed animal 900 lives in, for example, a farm cage 9. The farm cage 9 is made of metal or plastic capable of housing the farmed animal 900. The floor of the farm cage 9 is connected to a floor mesh 902, so the floor of the area where the farmed animal 900 defecates may be open. For example, a floor mesh 902 is installed inside the farm cage 9. The floor mesh 902 is made of metal or plastic. The floor mesh 902 is a plate with a mesh-like pattern of holes large enough for the feces of the farmed animal 900 to pass through but large enough for the farmed animal 900 itself to not pass through. In such a case, the farmed animal 900 lives on the floor mesh 902.

[0025] The floor mesh 902 is connected to the breeding cage 9 and the open floor surface. Unlike humans, animals may touch or eat feces. The floor mesh 902 is installed to keep the bedding 903 away from the animals. In some cases, learning is performed using the contour features of the feces to learn the feces property estimation model. For this reason, it is desirable to provide the floor mesh 902 in the breeding cage 9 where the shape of the feces is not distorted by the animals. In this way, the floor mesh 902 is installed so that the captive animals 900 do not touch the feces. However, if some effort is made to isolate the feces from the captive animals 900, the installation of the floor mesh 902 is not essential. Below, for the sake of simplicity, the excrement property estimation system 100 will be described taking the case where the floor mesh 902 is installed as an example.

[0026] A bedding 903 is placed under the floor mesh 902. The bedding 903 is preferably made of paper that can absorb moisture in order to clearly distinguish between diarrheal stool and urine.

[0027] When the farm animal 900 lives on the floor mesh 902, the feces of the farm animal 900 falls through the holes in the floor mesh 902 onto the bedding 903. Therefore, after the farm animal 900 excretes and before the feces are removed by cleaning, the feces are present on the bedding 903. Therefore, the space between the floor mesh 902 and the bedding 903 is an example of a feces space.

[0028] The photographing device 901 is installed at a position where it can photograph the feces space. Therefore, the photographing result image after the domestic animal 900 excretes and before the feces are removed by cleaning shows the feces of the domestic animal 900. The photographing device 901 is, for example, a camera. The photographing device 901 is installed at a position where it can photograph from the front so that the entire bedding 903 is captured. The photographing device 901 may be installed above the bedding 903 when the domestic animal 900 is kept in an environment where the domestic animal 900 does not come onto the feces even when photographing from above. The output of the photographing device 901 is, for example, a still image. The image quality of the still image output by the photographing device 901 is, for example, SD image quality (standard image quality) or higher. Hereinafter, for the sake of simplicity, the excrement property estimation system 100 will be described taking an example where the photographing device 901 is a camera with SD image quality.

[0029] The excrement property estimation system 100 uses a trained feces property estimation learning model (hereinafter referred to as the "feces property estimation model") to estimate whether or not abnormal feces (hereinafter referred to as "abnormal feces") is captured in the photographed image of the feces space to be analyzed. The feces space to be analyzed is a feces space in which feces from a farmed animal whose feces properties are to be estimated would be present if the farmed animal excretes feces. The feces property estimation model is an example of related information.

[0030] "Completed learning" means that a predetermined end condition (hereinafter referred to as "learning end condition") has been satisfied. The learning end condition is, for example, a condition that learning using a predetermined number of pieces of learning data has ended. The learning end condition may be, for example, a condition that the change due to the update of the fecal property estimation learning model is smaller than a predetermined change.

[0031] The excrement property estimation system 100 includes a feces property estimation model learning device 1 and a feces property estimation device 2. The feces property estimation model learning device 1 updates the feces property estimation learning model by machine learning using one or more pieces of learning data.

[0032] The learning data includes pairs of input-side learning data and correct answer labels. The input-side learning data includes learning image data. The learning image data is image data of a photographed image of a fecal space. In other words, the learning image data is image data of a photographed image of a fecal space. When multiple pieces of learning data are used in learning, images in which no feces are present may be used in learning as images shown by the learning image data. In other words, when multiple pieces of learning data are used in learning, it is not necessary for all of the multiple pieces of learning image data to be image data of images showing a fecal space in which feces is present. When multiple pieces of learning data are used in learning, some of the learning image data may be image data of images showing a fecal space in which no feces is present.

[0033] The correct answer label is information indicating whether or not abnormal stool is captured in the image of the corresponding learning image data. The correct answer label is a result of a pre-determination by a person who judges abnormal stool, such as an expert 910 who is skilled in distinguishing fecal properties, or a computer, as to whether or not abnormal stool is captured in the image of the learning image data.

[0034] The fecal property estimation device 2 estimates whether or not abnormal stool is captured in a photographed image of the fecal space to be analyzed (hereinafter referred to as the "analysis target image") using the fecal property estimation model obtained by the fecal property estimation model learning device 1. More specifically, the fecal property estimation device 2 acquires analysis target data including image data of the analysis target image (hereinafter referred to as the "analysis target image data"), and estimates the probability that abnormal stool is captured in the analysis target image using the fecal property estimation model obtained by the fecal property estimation model learning device 1.

[0035] Fig. 2 is a diagram showing a first example of a photographed image of a feces space in an embodiment. A floor mesh 902 and bedding 903 are shown in the image shown in Fig. 2. A plurality of feces are shown in area D1 in the image shown in Fig. 2. A plurality of food items are shown in area D2 in the image shown in Fig. 2. As shown in the image in Fig. 2, the feces and food are similar in color and shape. The space between the floor mesh 902 and bedding 903 in Fig. 2 is an example of a feces space.

[0036] 3 is a diagram showing a second example of a photographed image of a fecal space in an embodiment, in which abnormal feces is not shown in the photographed image of FIG.

[0037] 4 is a diagram showing a third example of a photographed image of a fecal space in an embodiment, in which no abnormal feces is shown in the photographed image of FIG.

[0038] 5 is a diagram showing a fourth example of a photographed image of a feces space in an embodiment. In the photographed image of FIG. 5, abnormal feces is photographed in an area D3.

[0039] 6 is a diagram showing a fifth example of a photographed image of a feces space in an embodiment, in which no abnormal feces is shown in the photographed image of FIG.

[0040] 7 is a diagram showing a sixth example of a photographed image of a feces space in the embodiment. In the photographed image of FIG. 7, abnormal feces are photographed in areas D4 and D5.

[0041] 8 is a diagram showing a seventh example of a photographed image of a feces space in the embodiment. In the photographed image of FIG. 8, abnormal feces is photographed in an area D6.

[0042] 9 is a diagram showing an eighth example of a photographed image of a feces space in an embodiment. The photographed image in FIG. 9 does not show abnormal feces.

[0043] 10 is a diagram showing a ninth example of a photographed image of a feces space in the embodiment. In the photographed image of FIG. 10, abnormal feces is photographed in an area D7.

[0044] 11 is a diagram showing an example of a hardware configuration of a fecal properties estimation model learning device 1 in an embodiment. The fecal properties estimation model learning device 1 includes a control unit 11 including a processor 91 such as a CPU (Central Processing Unit) and a memory 92 connected by a bus, and executes a program. The fecal properties estimation model learning device 1 functions as a device including the control unit 11, a communication unit 12, an input unit 13, a storage unit 14, and an output unit 15 by executing the program.

[0045] More specifically, in the fecal characteristics estimation model learning device 1, the processor 91 reads a program stored in the storage unit 14 and stores the read program in the memory 92. The processor 91 executes the program stored in the memory 92, causing the fecal characteristics estimation model learning device 1 to function as a device including a control unit 11, a communication unit 12, an input unit 13, a storage unit 14, and an output unit 15.

[0046] The control unit 11 controls the operation of each functional unit included in the fecal property estimation model learning device 1. The control unit 11, for example, learns the fecal property estimation learning model. The control unit 11, for example, records the results of the learning in the storage unit 14. The control unit 11, for example, controls the operation of the communication unit 12.

[0047] The communication unit 12 includes a communication interface for connecting the fecal property estimation model learning device 1 to an external device. The communication unit 12 communicates with the external device via a wired or wireless connection. The external device with which the communication unit 12 communicates is, for example, the fecal property estimation device 2. The communication unit 12 transmits a fecal property estimation model to, for example, the fecal property estimation device 2. The external device may be, for example, a device that transmits learning data. In such a case, the communication unit 12 receives learning data by communication with the device that transmits the learning data.

[0048] The input unit 13 includes input devices such as a mouse, a keyboard, and a touch panel. The input unit 13 may be configured as an interface that connects these input devices to the fecal characteristics estimation model learning device 1. The input unit 13 accepts input of various information to the fecal characteristics estimation model learning device 1. For example, learning data may be input to the input unit 13. An external storage device such as a Universal Serial Bus (USB) memory may be connected to the input unit 13.

[0049] The storage unit 14 is configured using a non-transitory computer-readable storage medium device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 14 stores various information related to the fecal property estimation model learning device 1. The storage unit 14 stores, for example, a program for controlling the operation of each functional unit of the fecal property estimation model learning device 1. The storage unit 14 stores, for example, a fecal property estimation learning model in advance. The storage unit 14 stores, for example, an updated fecal property estimation learning model.

[0050] Note that the learning data does not necessarily need to be input only to the communication unit 12, and does not necessarily need to be input only to the input unit 13. Each piece of information included in the learning data may be input from either the communication unit 12 or the input unit 13. For example, the input-side learning data may be input to the communication unit 12, and a correct answer label corresponding to the input-side learning data input to the communication unit 12 may be input to the input unit 13. Also, the learning data does not necessarily need to be acquired from the communication unit 12 or the input unit 13, and may be stored in advance in the storage unit 14.

[0051] The output unit 15 outputs various information. The output unit 15 includes a display device such as a CRT (Cathode Ray Tube) display, a liquid crystal display, or an organic EL (Electro-Luminescence) display. The output unit 15 may be configured as an interface that connects these display devices to the fecal property estimation model learning device 1. The output unit 15 outputs information input to the input unit 13, for example. The output unit 15 may display learning data input to the communication unit 12 or the input unit 13, for example. The output unit 15 may display the execution result of the fecal property estimation learning model, for example.

[0052] 12 is a diagram showing an example of the functional configuration of the control unit 11 in the embodiment. The control unit 11 includes a communication control unit 101, an input control unit 102, an output control unit 103, a learning data acquisition unit 104, a fecal property estimation learning model execution unit 105, an update unit 106, an end determination unit 107, and a recording unit 108.

[0053] The communication control unit 101 controls the operation of the communication unit 12. Under the control of the communication control unit 101, the communication unit 12 transmits the trained fecal property estimation learning model to the fecal property estimation device 2. The input control unit 102 controls the operation of the input unit 13. The output control unit 103 controls the operation of the output unit 15.

[0054] The learning data acquiring unit 104 acquires the learning data input to the communication unit 12 or the input unit 13. When the learning data has been recorded in advance in the storage unit 14, the learning data acquiring unit 104 may read the learning data from the storage unit 14.

[0055] The fecal property estimation learning model executing unit 105 executes learning of the fecal property estimation learning model on the input side learning data included in the learning data acquired by the learning data acquiring unit 104. By executing the fecal property estimation learning model, the fecal property estimation learning model executing unit 105 estimates the probability that abnormal stool is captured in the image indicated by the learning image data.

[0056] The update unit 106 updates the fecal property estimation learning model based on an estimation loss, which is the difference between the estimation result of the fecal property estimation learning model executing unit 105 and the correct label included in the learning data acquired by the learning data acquiring unit 104. The estimation result of the fecal property estimation learning model executing unit 105 is expressed by two components, for example, a component indicating the probability that abnormal stool is shown in the image indicated by the learning image data and a component indicating the probability that abnormal stool is not shown in the image indicated by the learning image data. The correct label is expressed by, for example, two components. One of the two components expressing the correct label is 1 when abnormal stool is shown in the image indicated by the learning image data and is 0 when abnormal stool is not shown in the image indicated by the learning image data. The other of the two components expressing the correct label is 0 when abnormal stool is shown in the image indicated by the learning image data and is 1 when abnormal stool is not shown in the image indicated by the learning image data.

[0057] When the estimation result of the fecal property estimation learning model executing unit 105 and the correct label are expressed by such two components, the estimation loss is expressed, for example, by an inner product of a two-component tensor indicating the estimation result of the fecal property estimation learning model executing unit 105 and a two-component tensor indicating the correct label. In this way, the estimation result of the fecal property estimation learning model executing unit 105 and the correct label are expressed, for example, by tensors. The tensor indicating the estimation result of the fecal property estimation learning model executing unit 105 and the tensor indicating the correct label may each be a first-order tensor (i.e., a vector), a second-order tensor (i.e., a matrix), or a third-order or higher tensor.

[0058] The end determination unit 107 determines whether or not a learning end condition is satisfied. The recording unit records various information in the storage unit .

[0059] 13 is a flowchart showing an example of the flow of processing executed by the fecal property estimation model learning device 1 according to the embodiment. Learning data is input to the communication unit 12 or the input unit 13, and the input learning data is acquired by the learning data acquisition unit 104 (step S101).

[0060] Next, the fecal property estimation learning model execution unit 105 executes the fecal property estimation learning model on the input side learning data included in the learning data acquired in step S101 (step S102). By executing the fecal property estimation learning model, the fecal property estimation learning model execution unit 105 estimates the probability that abnormal stool is present in the image indicated by the learning image data acquired in step S101.

[0061] Next, the update unit 106 updates the fecal property estimation learning model based on the difference between the estimation result obtained in step S102 and the correct label included in the learning data obtained in step S101 so as to reduce the difference (step S103).

[0062] Next, the end determination unit 107 determines whether or not the learning end condition is satisfied (step S104). If the learning end condition is satisfied (step S104: YES), the process ends. On the other hand, if the learning end condition is not satisfied (step S104: NO), the process returns to step S101.

[0063] 14 is a diagram showing an example of the hardware configuration of a fecal property estimation device 2 in an embodiment. The fecal property estimation device 2 includes a control unit 21 including a processor 93 such as a CPU and a memory 94 connected via a bus, and executes a program. The fecal property estimation device 2 functions as a device including the control unit 21, a communication unit 22, an input unit 23, a storage unit 24, and an output unit 25 by executing the program.

[0064] More specifically, in the fecal property estimation device 2, the processor 93 reads a program stored in the storage unit 24 and stores the read program in the memory 94. The processor 93 executes the program stored in the memory 94, causing the fecal property estimation device 2 to function as a device including a control unit 21, a communication unit 22, an input unit 23, a storage unit 24, and an output unit 25.

[0065] The control unit 21 controls the operation of each functional unit included in the fecal property estimation device 2. The control unit 21 executes, for example, a fecal property estimation model. The control unit 21 records, for example, the execution result of the fecal property estimation model in the storage unit 24. The control unit 21 controls, for example, the operation of the communication unit 22.

[0066] The communication unit 22 includes a communication interface for connecting the fecal property estimation device 2 to an external device. The communication unit 22 communicates with the external device via a wired or wireless connection. The external device with which the communication unit 22 communicates is, for example, the fecal property estimation model learning device 1. The communication unit 22 receives a fecal property estimation model from, for example, the fecal property estimation model learning device 1.

[0067] The external device may be, for example, a device that is a source of the analysis target data. In such a case, the communication unit 22 receives the analysis target data through communication with the device that is a source of the analysis target data.

[0068] The input unit 23 includes input devices such as a mouse, keyboard, and touch panel. The input unit 23 may be configured as an interface that connects these input devices to the fecal property estimation device 2. The input unit 23 accepts input of various information to the fecal property estimation device 2. For example, analysis target data may be input to the input unit 23. An external storage device such as a Universal Serial Bus (USB) memory may be connected to the input unit 23.

[0069] The storage unit 24 is configured using a non-transitory computer-readable storage medium device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 24 stores various information related to the fecal property estimation device 2. The storage unit 24 stores, for example, a program that controls the operation of each functional unit of the fecal property estimation device 2. The storage unit 24 stores, for example, a fecal property estimation model obtained in advance by the fecal property estimation model learning device 1. The storage unit 24 stores, for example, an estimation result based on the fecal property estimation model.

[0070] The analysis target data does not necessarily have to be input only to the communication unit 22, nor does it necessarily have to be input only to the input unit 23. The analysis target data may be input from either the communication unit 22 or the input unit 23. Moreover, the analysis target data does not necessarily have to be acquired from the communication unit 22 or the input unit 23, and may be stored in advance in the storage unit 24.

[0071] The output unit 25 outputs various information. The output unit 25 includes a display device such as a CRT display, a liquid crystal display, or an organic EL display. The output unit 25 may be configured as an interface that connects these display devices to the fecal property estimation device 2. The output unit 25 outputs information input to the input unit 23, for example. The output unit 25 may display analysis target data input to the communication unit 22 or the input unit 23, for example. The output unit 25 may display the execution result of a fecal property estimation model, for example. The output unit 25 outputs the estimation result of the fecal state and individual information, for example.

[0072] 15 is a diagram illustrating an example of a functional configuration of the control unit 21 in the embodiment. The control unit 21 includes a communication control unit 201, an input control unit 202, an output control unit 203, an analysis target data acquisition unit 204, an estimation unit 205, a recording unit 206, and an abnormality determination unit 207.

[0073] The communication control unit 201 controls the operation of the communication unit 22. The input control unit 202 controls the operation of the input unit 23. The output control unit 203 controls the operation of the output unit 25.

[0074] The analysis target data acquisition unit 204 acquires the analysis target data input to the communication unit 22 or the input unit 23. The analysis target data acquisition unit 204 may read out the analysis target data from the storage unit 24 when the analysis target data has been recorded in advance in the storage unit 24.

[0075] The estimation unit 205 executes the fecal property estimation model on the analysis target data acquired by the analysis target data acquisition unit 204. By executing the fecal property estimation model, the estimation unit 205 estimates the probability that abnormal stool is captured in the analysis target image included in the analysis target data.

[0076] The recording unit 206 records various types of information in the storage unit 24 .

[0077] The abnormality determination unit 207 determines whether the probability estimated by the estimation unit 205 is equal to or greater than a predetermined probability. When the probability estimated by the estimation unit 205 is equal to or greater than the predetermined probability, the abnormality determination unit 207 controls the operation of the communication unit 22 or the output unit 25 to notify a predetermined notification destination of a warning. The predetermined notification destination may be, for example, the output unit 25 itself. The predetermined notification destination may be, for example, a predetermined external device that is a communication destination of the communication unit 22.

[0078] FIG. 16 is a flowchart showing an example of the flow of processing executed by the fecal property estimation device 2 in the embodiment. The analysis target data acquisition unit 204 acquires the analysis target data (step S201). The processing of step S201 is executed, for example, when the photographing device 901 is started by a timer. It is preferable that the image quality of the analysis target image data obtained by the processing of step S201 is equivalent to that of the learning image data. Next, the estimation unit 205 estimates the probability that abnormal stool is captured in the analysis target image data acquired in step S201 by executing the fecal property estimation model (step S202). The estimation result is displayed, for example, on the output unit 25. The estimation result may be notified to a predetermined external device via, for example, the communication unit 22. For example, when the estimation result indicates that the probability of abnormal stool is estimated, the output unit 25 may display that the probability of abnormal stool is high, or the predetermined external device may be notified of the high probability of abnormal stool via the communication unit 22.

[0079] (Experimental Results) FIG. 17 is a diagram showing an example of an experiment result of estimating abnormal stool using the excrement property estimation system 100 of the embodiment. In the experiment, the captive animal 900 was a common marmoset. In the experiment, the learning data used included images obtained by the photographing device 901 and the results of an expert's judgment of whether the images were normal or abnormal. The learning data used in the experiment included 1,000 pieces of abnormal stool data and 10,000 pieces of normal stool data. In the experiment, a test was conducted to roughly classify images obtained on three days different from the day on which the learning data images were obtained into normal stool and abnormal stool. The data on the first day of the test data included 132 images of normal stool. The data on the first day of the test data included 15 images of abnormal stool. The data on the second day of the test data included 252 images of normal stool. The data on the second day of the test data included 29 images of abnormal stool. The data on the third day of the test data included 165 images of normal stool. The data on the third day of the test data included 13 images of abnormal stool. In the experiment, the accuracy rate was calculated for these data. The three days were October 22nd, October 30th, and November 4th.

[0080] FIG. 17 shows the results of estimating whether abnormal stool is present in the analysis target image data on October 22, October 30, and November 4 using multiple machine learning models, Model A, Model B, and Model C. Model A is a model that uses open source libraries Tensorflow and Keras, adds a fully connected layer to VGG16 in Keras, and performs transfer learning. Model B is a model that uses open source libraries Tensorflow and Keras, adds a fully connected layer to Xception in Keras, and performs transfer learning. Model C is a model that uses open source libraries Tensorflow and Keras, adds a fully connected layer to VGG19 in Keras, and performs transfer learning. FIG. 17 shows that the excrement property estimation system 100 succeeds in estimating abnormal stool with a high probability in any of Models A to C. Specifically, FIG. 17 shows that the cumulative accuracy rate is 85% or more in all three deep learning models. Fig. 17 shows that in model B, the accuracy rate is 90% or more for the data on October 22. Fig. 17 shows that in model C, the accuracy rate is 90% or more for the data on October 30.

[0081] In this way, the fecal characteristics estimation model learning device 1 obtains relationship information indicating the relationship between an image resulting from photographing the fecal space using learning data including the input learning data and the correct label and the probability that abnormal stool is captured in that image. The fecal characteristics estimation device 2 then uses the relationship information obtained by the fecal characteristics estimation model learning device 1 to estimate the probability that abnormal stool is captured in the image data to be analyzed.

[0082] The fecal property estimation model learning device 1 of the embodiment configured as described above obtains a fecal property estimation model through learning, and therefore, the fecal property estimation model learning device 1 can estimate the properties of animal feces with higher accuracy.

[0083] Furthermore, the excrement property estimation system 100 of the embodiment configured as described above includes a feces property estimation device 2 that executes the feces property estimation model obtained by the feces property estimation model learning device 1. Therefore, the excrement property estimation system 100 can estimate the properties of animal feces with higher accuracy.

[0084] Furthermore, the excrement property estimation system 100 of the embodiment configured in this manner estimates the feces of a farmed animal 900 kept in a breeding cage 9 equipped with a floor mesh 902. Due to the presence of the floor mesh 902 as described above, the probability that the farmed animal 900 will tamper with the feces is low, and the shape of the feces is less likely to be distorted. This makes it possible to obtain a feces property estimation model with high estimation accuracy. As a result, the excrement property estimation system 100 is able to estimate the properties of the animal's feces with higher accuracy.

[0085] In addition to feces and urine, various other things such as food, sawdust, and pieces of the food box fall onto the flooring 903 of the breeding cage 9. In particular, the area and shape of food are very similar to normal feces. The position where the feces falls is not uniform, but falls randomly. Furthermore, the shape of abnormal feces also changes organically in various ways. The fecal property estimation model learning device 1 uses an image captured by the imaging device 901, such as an image of the entire flooring 903, that includes not only an image of the feces but also information about the surroundings of the feces, for learning. Therefore, the fecal property estimation model learning device 1 can obtain a fecal property estimation model that is less influenced by information other than the feces and more influenced by information about the feces during estimation. Therefore, the fecal property estimation model learning device 1 makes it possible to estimate the properties of the animal's feces with higher accuracy.

[0086] As described above, the photographing device 901 is not a special or expensive camera but a camera with SD standard image quality. Therefore, the excrement property estimation system 100 can estimate the condition of the stool using images taken with a general camera with SD standard image quality, without using a special or expensive camera. Note that the excrement property estimation system 100 can also estimate the condition of the stool using monochrome images.

[0087] The excrement property estimation system 100 enables the detection of abnormal feces, which is one of the indicators of animal health, to be performed automatically using a computer rather than manually. Therefore, the excrement property estimation system 100 enables labor-saving in the health management of animals in animal breeding facilities.

[0088] (Modification) The excrement property estimation system 100 may include a feces property estimation device 2a instead of the feces property estimation device 2. The feces property estimation device 2a determines the properties of excrement for each of the multiple breeding cages 9.

[0089] Fig. 18 is a diagram showing an example of the hardware configuration of a fecal property estimation device 2a in a modified example. Hereinafter, components having the same functions as those in the fecal property estimation device 2 are given the same reference numerals as in Fig. 14 and will not be described. The fecal property estimation device 2a includes a control unit 21a including a processor 95 such as a CPU and memory 96 connected by a bus, and executes a program. The fecal property estimation device 2a functions as a device including the control unit 21a, communication unit 22a, input unit 23, storage unit 24 and output unit 25 by executing the program.

[0090] More specifically, in the fecal property estimation device 2a, the processor 95 reads a program stored in the storage unit 24 and stores the read program in the memory 96. When the processor 95 executes the program stored in the memory 96, the fecal property estimation device 2a functions as a device including a control unit 21a, a communication unit 22a, an input unit 23, a storage unit 24, and an output unit 25.

[0091] The control unit 21a controls the operation of each functional unit included in the fecal property estimation device 2a. The control unit 21a executes a fecal property estimation model, for example, for each of the multiple breeding cages 9. The control unit 21a records, for example, the execution result of the fecal property estimation model in the storage unit 24. The control unit 21a controls, for example, the operation of the communication unit 22a.

[0092] The communication unit 22a includes a communication interface for connecting the fecal property estimation device 2a to an external device. The communication unit 22a communicates with the external device via wired or wireless communication. The external device with which the communication unit 22a communicates is, for example, the fecal property estimation model learning device 1. The communication unit 22a acquires a fecal property estimation model from, for example, the fecal property estimation model learning device 1.

[0093] The external device is, for example, a photographing device 901 installed in each of the multiple breeding cages 9. The communication unit 22a receives analysis target data for each breeding cage 9 through communication with each photographing device 901.

[0094] Fig. 19 is a diagram showing an example of a functional configuration of a control unit 21a in a modified example. Hereinafter, components having the same functions as those in the control unit 21 are denoted by the same reference numerals as those in Fig. 15, and description thereof will be omitted. The control unit 21a includes a communication control unit 201a, an input control unit 202, an output control unit 203, an analysis target data acquisition unit 204a, an estimation unit 205a, a recording unit 206, and an abnormality determination unit 207a.

[0095] The communication control unit 201a controls the operation of the communication unit 22a.

[0096] The analysis target data acquisition unit 204a acquires the analysis target data of each breeding cage 9 input to the communication unit 22a.

[0097] The estimation unit 205a executes a feces property estimation model for each breeding cage 9 on the analysis target data of each breeding cage 9 acquired by the analysis target data acquisition unit 204a. By executing the feces property estimation model, the estimation unit 205a estimates the probability that abnormal stool is captured in the analysis target image included in the analysis target data for each breeding cage 9.

[0098] The abnormality determination unit 207a determines for each breeding cage 9 whether the probability estimated by the estimation unit 205a is equal to or greater than a predetermined probability. When an abnormal breeding cage is present, the abnormality determination unit 207a controls the operation of the communication unit 22 or the output unit 25 to notify a predetermined notification destination of information indicating the abnormal breeding cage. An abnormal breeding cage is a breeding cage 9 for which the probability estimated by the estimation unit 205a is equal to or greater than a predetermined probability. The predetermined notification destination may be, for example, the output unit 25 itself. The predetermined notification destination may be, for example, a predetermined external device that is a communication destination of the communication unit 22.

[0099] In this way, the excrement property estimation system 100 equipped with the feces property estimation device 2a instead of the feces property estimation device 2 notifies the abnormal breeding cages among the multiple breeding cages 9. This allows the breeding manager to arrive at the abnormal breeding cage at an earlier time, and the abnormal breeding animal can receive care at an earlier stage. This prevents the breeding animal from becoming seriously ill. As a result, in a breeding facility having multiple breeding cages 9, the effort required for the breeding manager to check the breeding cages 9 can be reduced.

[0100] It is not necessary for the excrement property estimation system 100 to estimate whether the stool is abnormal or non-abnormal. During learning by the feces property estimation model learning device 1, learning data with more detailed annotations such as no stool, loose stool, muddy stool, watery stool, etc. is used, thereby obtaining a feces property estimation model that classifies the state of the stool in the input image in more detail and estimates the probability of each classification. Therefore, by using the feces property estimation model obtained in this manner, the excrement property estimation system 100 can classify the input image in more detail and estimate the probability of each classification.

[0101] In addition, in learning, image data of an extracted stool image in which a stool portion is extracted from a photographed image in advance may be used as learning image data. The stool portion is a portion obtained by enclosing a stool area on a screen with a rectangle having a minimum area among rectangles that satisfy the condition that the stool area can be enclosed. The extracted stool image is an image of the extracted stool. The extracted stool is stool having a horizontal projection area of ​​a predetermined area or more. That is, the learning image data may be image data of an extracted stool image, which is an image of stool that is extracted in advance from a photographed image and has a horizontal projection area of ​​a predetermined area or more. In such a case, the stool property estimation model learning device 1 learns the stool property estimation learning model using image data of an image that does not include an obstructing object. Therefore, a mathematical model that has a higher accuracy of distinguishing between stool and obstructing objects is obtained as the stool property estimation model. The obstructing object is an object other than stool that is close to stool in color and shape within a predetermined reference value. When the subject of property estimation by excrement property estimation system 100 is feces, the obstructing object is something other than feces, whereas when the subject of property estimation by excrement property estimation system 100 is excrement, the obstructing object is something other than excrement. In other words, the obstructing object is something other than excrement, the subject of property estimation by excrement property estimation system 100.

[0102] In addition, when learning, image data of an image obtained by performing image processing on the captured image in advance using a filter that does not pass the color of obstacles such as food may be used as learning image data. In such a case, the fecal property estimation model learning device 1 learns the fecal property estimation learning model using image data of an image that does not include obstacles. Therefore, a mathematical model that has a higher accuracy of distinguishing between feces and obstacles is obtained as the fecal property estimation model.

[0103] In addition to the learning image data, the input-side learning data may further include information (hereinafter referred to as "photographing date and time information") indicating the date and time when the learning image data was photographed (hereinafter referred to as "photographing date and time"). In such a case, the learning (i.e., updating) of the fecal properties estimation learning model is performed based on not only the learning image data but also the photographing date and time. That is, when learning the fecal properties estimation learning model, an estimation of the probability that abnormal stool is captured in the image shown by the learning image data is performed based on the learning image data and the photographing date and time, and the learning of the fecal properties estimation learning model is updated based on the estimation result.

[0104] Therefore, in such a case, the feces characteristics estimation model obtained by the feces characteristics estimation model learning device 1 is a mathematical model that estimates the probability that abnormal stool is captured in the analysis target image data based on the analysis target image data and the photographing date and time when the analysis target image data was captured. Therefore, in such a case, the analysis target data further includes, in addition to the analysis target image data, photographing date and time information indicating the photographing date and time when the analysis target image data was captured.

[0105] The effect of including the photographing date and time information will be described. Animals may have the habit of behaving according to the date and time, such as being nocturnal. Therefore, the state of the feces shown in the photographed image may have a correlation with the date and time. Therefore, by performing learning of the fecal property estimation learning model based not only on the learning image data but also on the photographing date and time, a mathematical model that performs estimation using information on the animal's habits according to the date and time when estimating the probability of the presence or absence of abnormal stool is obtained as the fecal property estimation model. Therefore, by performing learning using the photographing date and time information as well, the accuracy of estimation by the fecal property estimation device 2 is further improved.

[0106] In particular, many diurnal animals (sleeping at night) such as marmosets tend to defecate early in the morning. For animals that defecate early in the morning, it is desirable to photograph the state of the bedding first thing in the morning and obtain an image of the feces in a fresh state. This is because fresh feces are the easiest to determine the state of the feces. In this way, the photographing date and time information is related to the state of the feces. Many other animals also have habits regarding the time of defecation. Therefore, the estimation accuracy is improved by obtaining image data at the time when fresh feces are produced. Conversely, if a large amount of feces is produced at a time that is different from the habit, there is a high probability that the captive animal is in an abnormal state and has a high possibility of producing abnormal feces. Therefore, the photographing date and time information is related to the state of the feces. When making a daily judgment, the estimation accuracy is improved by using the information from the previous day.

[0107] It is not necessary that only one animal is kept in the breeding cage 9. When multiple animals are kept in the breeding cage 9, feces of multiple animals will be present on the bedding 903.

[0108] In learning, information indicating the maximum number of animals from which feces exist in the feces space (hereinafter referred to as "number information") may be used. Specifically, the input-side learning data may include the number information.

[0109] Note that loose stool and vomit look similar and are difficult to distinguish from each other. However, both indicate abnormalities in farmed animals. Therefore, the excrement property estimation system 100 may also determine that vomit is abnormal stool.

[0110] The amount of feces is approximately proportional to the number of animals kept in the breeding cage 9. And, all of the animals kept in the breeding cage 9 do not necessarily produce abnormal feces at the same time. Therefore, as the number of animals kept in the breeding cage 9 increases, the frequency of the ratio of abnormal feces being lower than normal feces, which is normal feces, increases. This means that as the number of animals kept in the breeding cage 9 increases, it becomes more difficult to estimate whether or not abnormal feces are present. Therefore, by performing learning using the number information as well, the fecal property estimation model learning device 1 can obtain a mathematical model that also learns the relationship between the number of animals and abnormal feces as a fecal property estimation model. Therefore, by performing learning using the number information as well, the accuracy of estimation by the fecal property estimation device 2 and the fecal property estimation device 2a is further improved. Note that, when the number information is also used in learning, the analysis target data also includes the number information. The number information included in the analysis target data indicates the number of animals to be estimated.

[0111] When learning, information indicating the type of food given to the captive animals at a predetermined timing before the time of photographing, such as immediately before the time of photographing, (hereinafter referred to as "food information") may be used. Specifically, the input side learning data may include the food information. Note that the time of photographing is the time when photographing is performed by the photographing device 901 to obtain the photographed image.

[0112] Since feces is formed by digesting food, there is a correlation between food and feces. Thus, by performing learning using food information as well, the fecal property estimation model learning device 1 can obtain a mathematical model that also learns the relationship between food and abnormal stool as a fecal property estimation model. Therefore, by performing learning using food information as well, the accuracy of estimation by the fecal property estimation device 2 and the fecal property estimation device 2a is further improved. Note that when food information is used during learning, the data to be analyzed also includes the food information.

[0113] As mentioned above, food has a correlation with stool. Therefore, the accuracy of the estimation can be improved by including information about the food. Normal stool often has a similar shape regardless of the food. However, abnormal stool may contain undigested matter, and the influence of the food is likely to be evident in the condition of the stool. Therefore, the accuracy of the estimation can be improved by using information about the type of food given (color, ingredients, manufacturer, etc. of the food). Furthermore, the condition of the food may be included in the information about the food.

[0114] Furthermore, if vomit falls onto the bedding, it is difficult to distinguish it from feces. However, since the influence of food is stronger in vomit than in feces, the accuracy of distinguishing between feces and vomit can be improved by using information on food. As a result, the accuracy of estimating vomit by excrement property estimation system 100 is improved. Furthermore, as described above, vomit also indicates abnormalities in the farmed animal. Therefore, excrement property estimation system 100 may determine that there is an abnormality in the farmed animal even when the presence of vomit is estimated.

[0115] When learning, information about domestic animals (hereinafter referred to as "domestic animal information") may be used. Specifically, the input-side learning data may include domestic animal information. The domestic animal information may include information about the domestic animal's taxonomic species, genetic lineage, sex, maturity (age in days, weeks, months, or years), etc.

[0116] Fecal properties vary depending on the characteristics of the farmed animals, such as species, genetic lineage, sex, and maturity. Thus, by performing learning using farmed animal information as well, the fecal property estimation model learning device 1 can obtain a mathematical model that also learns the relationship between the farmed animal's characteristics and abnormal feces as a fecal property estimation model. Therefore, by performing learning using farmed animal information as well, the accuracy of estimation by the fecal property estimation device 2 and the fecal property estimation device 2a is further improved. Note that when farmed animal information is also used in learning, feed information is also included in the analysis target data.

[0117] The image capturing device 901 may be a camera that captures images using visible light, or may be an infrared camera. The captured image output by the image capturing device 901 may be either monochrome or color.

[0118] The communication units 12, 22, and 22a may be configured to include an interface for connecting to an external storage device such as a Universal Serial Bus (USB) memory that stores various information. In this case, the communication units 12, 22, and 22a may output information to the connected external storage device.

[0119] The fecal property estimation model learning device 1 and the fecal property estimation device 2 do not necessarily need to be implemented as separate devices. For example, the fecal property estimation model learning device 1 and the fecal property estimation device 2 may be implemented as a single device having both functions.

[0120] The fecal property estimation model learning device 1 and the fecal property estimation device 2a do not necessarily need to be implemented as separate devices. For example, the fecal property estimation model learning device 1 and the fecal property estimation device 2a may be implemented as a single device having both functions.

[0121] The excrement property estimation system 100 may be implemented as a system using a plurality of information processing devices communicatively connected via a network. In this case, the functional units of the feces property estimation model learning device 1 and the feces property estimation device 2 may be distributed and implemented in the plurality of information processing devices. Also, the functional units of the feces property estimation model learning device 1 and the feces property estimation device 2a may be distributed and implemented in the plurality of information processing devices.

[0122] All or part of the functions of the fecal property estimation model learning device 1, the fecal property estimation device 2, and the fecal property estimation device 2a may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into a computer system. The program may be transmitted via an electric communication line.

[0123] The feces space is an example of an excrement space. Abnormal stool is an example of abnormal excrement. The feces property estimation learning model is an example of an excrement property estimation learning model. The feces property estimation learning model execution unit 105 is an example of an excrement property estimation learning model execution unit. The feces property estimation model learning device 1 is an example of an excrement property estimation model learning device. The feces property estimation devices 2 and 2a are examples of excrement property estimation devices. The extracted feces image is an example of an extracted excrement image.

[0124] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Explanation of symbols]

[0125] 100...Excrement property estimation system, 1...Feces property estimation model learning device, 2, 2a...Feces property estimation device, 11...Control unit, 12...Communication unit, 13...Input unit, 14...Memory unit, 15...Output unit, 101...Communication control unit, 102...Input control unit, 103...Output control unit, 104...Learning data acquisition unit, 105...Feces property estimation learning model execution unit, 106...Update unit, 107...End determination unit, 108...Recording unit, 21, 21a...Control unit, 22, 22a...Communication unit, 23...Input unit, 24...Memory unit, 25...Output unit, 201, 201a...Communication control unit, 202...Input control unit, 203...Output control unit, 204, 204a...Analysis target data acquisition unit, 205, 205a...Estimation unit, 206: recording unit; 207, 207a: abnormality determination unit; 91: processor; 92: memory; 93: processor; 94: memory; 95: processor; 96: memory; 900: captive animal; 901: imaging device; 902: floor mesh; 903: floor mat; 910: expert

Claims

1. a learning data acquisition unit that acquires learning data including an excrement space, which is a predetermined space in which excrement from a bred animal exists, input side learning data including learning image data, which is image data of an image obtained by photographing the excrement space, and a correct answer label indicating whether or not an abnormal excrement is captured in an image represented by the learning image data; an excrement property estimation learning model execution unit that estimates the probability that abnormal excrement is shown in an image represented by the learning image data, using an excrement property estimation learning model that is a machine learning model that estimates the probability that abnormal excrement is shown in the learning image data; an update unit that updates the excrement property estimation learning model based on a difference between an estimation result of the excrement property estimation learning model execution unit and the correct label so as to reduce the difference; An excrement property estimation model learning device comprising:

2. The learning image data is an image extracted in advance from a photographed image, and is image data of an extracted excrement image, which is an image of excrement having a horizontal projection area equal to or larger than a predetermined area. The excrement property estimation model learning device according to claim 1 .

3. The learning image data is image data of an image obtained by performing image processing on the photographed image in advance using a filter that does not pass the colors of obstacles other than excrement that are similar in color and shape to excrement within a predetermined reference value. The excrement property estimation model learning device according to claim 1 or 2.

4. The input side learning data includes information indicating a date and time when the learning image data was captured. The excrement property estimation model learning device according to claim 1 .

5. the input-side learning data includes information indicating a maximum number of animals from which excrement exists in the excrement space; The excrement property estimation model learning device according to claim 1 .

6. the input-side learning data includes information indicating food given to the animal to be raised at a predetermined timing before a time when an image is captured by an image capturing device to obtain a captured image; The excrement property estimation model learning device according to claim 1 .

7. an analysis target data acquisition unit that acquires analysis target data including an excrement space, which is a predetermined space in which excrement of a bred animal exists, and image data of a photographed image of the excrement space; an estimation unit that estimates the probability that abnormal excrement is captured in the image data acquired by the analysis target data acquisition unit using a trained excrement property estimation learning model obtained by an excrement property estimation model learning device comprising: a learning data acquisition unit that acquires learning data including input-side learning data including learning image data, which is image data of a photographed image, and a correct answer label that indicates whether or not an abnormal excrement, which is an abnormal excrement, is captured in the image represented by the learning image data; an excrement property estimation learning model execution unit that estimates the probability that abnormal excrement is captured in the image represented by the learning image data using an excrement property estimation learning model, which is a machine learning model that estimates the probability that abnormal excrement is captured in the learning image data; and an update unit that updates the excrement property estimation learning model based on a difference between an estimation result of the excrement property estimation learning model execution unit and the correct answer label so as to reduce the difference; An excrement property estimation device comprising:

8. a learning data acquisition step of acquiring learning data including an excrement space, which is a predetermined space in which excrement of a bred animal exists, input side learning data including learning image data, which is image data of an image obtained by photographing the excrement space, and a correct answer label indicating whether or not an abnormal excrement is captured in the image shown by the learning image data; an excrement property estimation learning model execution step of estimating the probability that abnormal excrement is shown in the image represented by the learning image data, using an excrement property estimation learning model which is a machine learning model that estimates the probability that abnormal excrement is shown in the learning image data; an updating step of updating the excrement property estimation learning model based on a difference between the estimation result of the excrement property estimation learning model executing step and the correct label so as to reduce the difference; The excrement property estimation model learning method includes the steps of:

9. an analysis target data acquisition step of acquiring analysis target data including an excrement space, which is a predetermined space in which excrement of a bred animal exists, and image data of a photographed image of the excrement space; an estimation step of estimating the probability that abnormal excrement is captured in the image data obtained in the analysis target data acquisition step, using a trained excrement property estimation learning model obtained by an excrement property estimation model learning device including: a learning data acquisition unit that acquires learning data including input side learning data including learning image data, which is image data of a photographed image, and a correct answer label that indicates whether or not an abnormal excrement, which is an abnormal excrement, is captured in the image represented by the learning image data; an excrement property estimation learning model execution unit that estimates the probability that abnormal excrement is captured in the learning image data, using an excrement property estimation learning model that is a machine learning model that estimates the probability that abnormal excrement is captured in the learning image data; and an update unit that updates the excrement property estimation learning model based on a difference between an estimation result of the excrement property estimation learning model execution unit and the correct answer label so as to reduce the difference; The method for estimating excrement properties comprises the steps of:

10. A program for causing a computer to function as the excrement property estimation model learning device according to any one of claims 1 to 6.

11. A program for causing a computer to function as the excrement property estimation device according to claim 7.

Images