Weight measurement system and weight measurement method

Abstract

To provide a body weight measuring system capable of measuring a body weight of an object animal appropriately.SOLUTION: A body weight measuring system which images an object animal with a camera 26 includes a machine learning model 41 with the plane dimension and width information 33 on the object animal and the side surface dimension and width information 34 on the object animal as explanatory variables, and with weight information on the object animal as objective variables, and a calibration part 60 for performing calibration of the machine learning model 41.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] This invention relates to a body weight measurement system and a body weight measurement method.

Background Art

[0002] The body weight measurement device for pigs described in Patent Document 1 attaches a dust-proof box to the ceiling of a pigsty near a water feeder, and installs a projector equipped with a grid stripe slide and a video camera inside this dust-proof box. When light is projected from the projector, grid stripes are projected onto the floor and the body surface of the pig. This image is captured by the video camera, and based on the deviation of the grid stripes in the captured image, the body height (H) of the pig is determined. Also, the captured image is subjected to binary black and white processing to obtain the projected area (A) of the pig. There is a close relationship of W = aH^bA^c between the body height (H), projected area (A), and body weight (W) of the pig. Furthermore, since the dimension (d = b + 2c) of its multiple regression formula is close to the dimension of volume (d = 3), the body weight (W) is calculated using the body height (H) and the projected area (A).

[0003] Patent Document 2 discloses that during the fish farming process or at the time of shipping, three selection criteria of only body length, body length and body weight, and only body weight are provided, and by selecting the criteria, the body length and body weight can be accurately measured efficiently without weakening the fish while alive, and a measurement device that can perform multi-stage selection is provided.

[0004] This measurement device provides a weighing chute that slopes downward to flow a certain amount of water, and a passing detector for detecting fish passing through this weighing chute is attached in the middle. A light source and a camera capable of taking a still image of the fish are provided above and below the weighing chute according to the detection signal of this detector. On the other hand, the movable part of a load cell that transmits an output proportional to the body weight of the fish is attached to the weighing chute, and the fixed part is attached to the weighing instrument frame respectively. Image processing is performed by an image processing device and an image memory built into the control device, and this image and the load cell output are respectively calculated into body length and body weight data by the CPU in the control device.The selection criterion setting device of the control device compares with the three selection criteria of only input body length, body length and body weight, and only body weight, and is configured to perform multi-stage sorting with a sorting device provided in the subsequent stage.

[0005] Patent Document 3 describes the provision of an animal weight measuring device, a measuring facility, and an automatic measuring device.

[0006] This weight measuring device consists of a load sensor, a measuring platform mounted horizontally above it and having a predetermined size according to the target animal, side fences positioned approximately vertically at the left and right ends of the measuring platform and having a predetermined height and depth according to the target animal, and a front end fence installed on the front end of the measuring platform and lower in height than the side fences. The rear end between the side fences is left open. The measurement facility consists of one or more weight measuring devices arranged in a ring or horizontal row, a feeder 11 positioned near the front of the front end fence, and a barrier member to prevent entry to the feeder side positioned outside the side fences. The automatic measurement device consists of a weight measuring device, a cable for transmitting its sensor signal, and a data acquisition and processing device. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2002-286421 [Patent Document 2] Japanese Patent Application Publication No. 8-050052 [Patent Document 3] Japanese Patent Publication No. 2003-114145 [Overview of the project] [Problems that the invention aims to solve]

[0008] In the method described in Patent Document 1, it is difficult to project a grid pattern onto the floor and the surface of the pig's body, and accurate measurements are not always possible. In Patent Document 2, the method is based on measuring fish and has the problem of being difficult to apply to measuring the weight of other animals.

[0009] Patent Document 3 requires guiding the animals to the weighing platform, which is difficult. The biggest challenge for pig farms is the aging workforce and the declining working population. Weighing pigs requires a great deal of effort, which means that they cannot weigh the pigs and are not able to ship them at the optimal time.

[0010] Furthermore, as shown in Figure 10(A), even when pigs of the same age are shipped, there are individual differences in growth even within the same pen, resulting in variations in weight as shown in Figure 10(B). The variation in meat grade after shipment is as shown in Figure 11 on a national average. The portion with a high unit price enclosed by the box W in Figure 11 accounts for slightly over 50%, and the challenge is how to increase this portion.

[0011] For example, as shown in Figure 12(A), statistics on the weight of pigs raised on a certain farm and shipped on a certain day show that less than 50% of the pigs are of the "superior" grade, weighing between 70kg and 80kg, missing the optimal point. If the optimal shipping date is predicted so that more than 65% of the pigs are of the "superior" grade, as shown in Figure 12(B), it is estimated that a profit of 450 yen per pig can be earned, which is the price difference between Figure 12(A) and Figure 12(B). Furthermore, if it is possible to predict the date and time when more than 65% of the pigs will be of the "superior" grade, as shown in Figure 12(B), it is possible to reduce feed costs. For example, reducing feed costs for 8 days can save 1600 yen per pig.

[0012] Embodiments of the present invention have been made in view of the above points, and aim to provide a weight measurement system capable of appropriately measuring the weight of a target animal. [Means for solving the problem]

[0013] The weight measurement system according to an embodiment of the present invention includes: an image acquisition means that captures images of a target animal with a camera and obtains a planar image from above and a lateral image of the target animal; a planar dimension / width information acquisition means that obtains planar dimension / width information as planar width information and planar length information of the target animal based on the planar image; a lateral dimension / width information acquisition means that obtains lateral dimension / width information of the target animal based on the lateral image; a machine learning model obtained by machine learning with the planar dimension / width information and the lateral dimension / width information as explanatory variables and the weight information of the target animal as the objective variable; and a weight information acquisition means that provides the obtained planar dimension / width information and the lateral dimension / width information as explanatory variables to the machine learning model to obtain the weight information, which is the objective variable. A calibration unit that performs calibration of the machine learning model; an individual information management means that uses the image obtained by the image acquisition means to perform individual identification of the target animal, attaches identification information to the image of each individual target animal, and provides the image with the attached identification information to the planar dimension / area information acquisition means, the side dimension / area information acquisition means and the weight information acquisition means; an appropriate image selection means that, when providing the image with the attached identification information to the planar dimension / area information acquisition means, the side dimension / area information acquisition means and the weight information acquisition means, determines the posture of the target animal by detecting the curvature of the spine, and performs processing to select only the images that can be processed appropriately based on this determination result; and the appropriate image selection means Planar image The system comprises a correction processing means that performs an elliptical approximation along the body interface of the target animal, calculates the angle of the ellipse, and performs a correction process that rotates the image using an affine transformation with the center of the ellipse as the axis of rotation so that the angle of the ellipse becomes horizontal. The weight measurement system further comprises a deviation data acquisition means that performs elliptical approximation along the body boundary of the target animal on a planar image selected by the appropriate image selection means, and obtains deviation data between the length of the center line of the obtained ellipse and the length of the line segment obtained by connecting the highest points in each divided area obtained by dividing the planar image into multiple vertical sections. The machine learning model is obtained by machine learning with the planar dimensions / area information and the side dimensions / area information as explanatory variables, the deviation data obtained by the deviation data acquisition means as explanatory variables, and the weight information of the target animal as the target variable. The weight information acquisition means adds the deviation data obtained by the deviation data acquisition means as an explanatory variable to the planar dimensions / area information and the side dimensions / area information as explanatory variables and provides it to the machine learning model to obtain the weight information, which is the target variable. It is characterized by the following:

[0014] The weight measurement method according to an embodiment of the present invention includes: an image acquisition step of capturing an image of a target animal with a camera and obtaining a planar image from above and a lateral image of the target animal; a planar dimension and width information acquisition step of obtaining planar dimension and width information as planar width information and planar length information of the target animal based on the planar image; a lateral dimension and width information acquisition step of obtaining lateral dimension and width information of the target animal based on the lateral image; a weight information acquisition step of using a machine learning model obtained by machine learning with the planar dimension and width information and the lateral dimension and width information as explanatory variables and the weight information of the target animal as the target variable, and providing the obtained planar dimension and width information and the lateral dimension and width information as explanatory variables to the machine learning model to obtain weight information, which is the target variable; a history information storage step of storing pairs of explanatory variables and target variables as history information when weight information is acquired using the machine learning model in the weight information acquisition step; and the explanatory variables in the pairs in the history information stored in the history information storage step. The process includes: an updated learning model generation step, which obtains an updated machine learning model by performing the same machine learning as when the machine learning model was created, using the actual weight information of the corresponding target animal as an explanatory variable and the actual weight information as the target variable; an individual information management step, which performs individual identification of the target animal using the images obtained in the image acquisition step, attaches identification information to the image of each individual target animal, and provides the images with the attached identification information to the planar dimension / area information acquisition step, the side dimension / area information acquisition step, and the weight information acquisition step; an appropriate image selection step, which determines the posture of the target animal by detecting the curvature of the spine when providing the images with the attached identification information to the planar dimension / area information acquisition step, the side dimension / area information acquisition step, and the weight information acquisition step, and selects only the images that can be processed appropriately based on this determination result; and the selection made in the appropriate image selection step. Planar imageAn elliptical approximation is performed along the boundary surface of the body of the target animal, the angle of the ellipse is calculated, and a correction process is performed to rotate the image using an affine transformation with the center of the ellipse as the rotation axis so that the angle of the ellipse that levels the image becomes horizontal. The correction process step is provided. A method for measuring weight, The process further includes a deviation data acquisition step, which involves performing an elliptical approximation along the body boundary of the target animal to the planar image selected in the appropriate image selection step, and obtaining deviation data between the length of the center line of the obtained ellipse and the length of the line segment obtained by connecting the highest points in each divided area obtained by dividing the planar image vertically into multiple sections. The machine learning model is obtained by machine learning with the planar dimensions / area information and the side dimensions / area information as explanatory variables, the deviation data obtained in the deviation data acquisition step as explanatory variables, and the target animal's weight information as the target variable. The weight information acquisition step involves adding the deviation data obtained in the deviation data acquisition step as explanatory variables to the planar dimensions / area information and the side dimensions / area information as explanatory variables, and providing this to the machine learning model to obtain the target variable, weight information. It is characterized by this.

Brief Explanation of Drawings

[0015] [Figure 1] Configuration diagram when the weight measurement system according to the embodiment of the present invention is configured by a computer. [Figure 2] Functional block diagram of the weight measurement system according to the first embodiment of the present invention. [Figure 3] Explanation diagram of the planar dimension / width information acquired by the weight measurement system according to the embodiment of the present invention. [Figure 4] Explanation diagram of the side dimension / width information acquired by the weight measurement system according to the embodiment of the present invention. [Figure 5] Flowchart showing the operation of the weight measurement system according to the first embodiment of the present invention. [Figure 6] Functional block diagram of the weight measurement system according to the second embodiment of the present invention. [Figure 7] Explanation diagram of proper image selection by the weight measurement system according to the embodiment of the present invention. [Figure 8] Explanation diagram of correction processing by the weight measurement system according to the embodiment of the present invention. [Figure 9] Flowchart showing the operation of the weight measurement system according to the second embodiment of the present invention. [Figure 10] Diagram showing the variation in weight due to the growth of pigs, which are the target animals in the embodiments of the present invention. [Figure 11] Diagram showing the variation in weight when the pigs, which are the target animals in the embodiments of the present invention, are refined. [Figure 12]A diagram showing the weight variation of pigs at the time of shipment and the ideal weight distribution, which are the target animals of the embodiment of the present invention. [Figure 13] A figure showing the estimation accuracy of a weight measurement system according to an embodiment of the present invention at a certain point in time. [Modes for carrying out the invention]

[0016] Hereinafter, a weight measurement system and weight measurement method according to an embodiment of the present invention will be described with reference to the attached drawings. In each figure, the same components are denoted by the same reference numerals, and redundant explanations are omitted. Figure 1 shows a computer configuration diagram when the weight measurement system according to an embodiment of the present invention is configured using a computer. That is, the CPU 10 configures the weight measurement system using programs and data in the main memory 11. An external storage interface 13, an input interface 14, a display interface 15, and a data input interface 16 are connected to the CPU 10 via a bus 12.

[0017] An external storage device 23 is connected to the external storage interface 13. The external storage device 23 stores programs and data necessary for the operation of this weight measurement system, which the CPU 10 can read and use from the main memory 11 as needed. For this reason, the external storage device 23 stores programs that implement the machine learning models and various means described later. An input device 24 such as a keyboard or touch panel and a pointing device 22 such as a mouse are connected to the input interface 14. A display device 25 having a screen such as an LCD is connected to the display interface 15. Cameras 26-1 to 26-m are connected to the data input interface 16, and these cameras 26-1 to 26-m function to image the target animal. The number of cameras 26-1 to 26-m is arbitrary. Image data obtained by cameras 26-1 to 26-m is taken in by the data input interface 16 to the CPU 10, display interface 15, and other necessary parts.

[0018] Figure 2 shows a functional block diagram of a weight measurement system according to the first embodiment of the present invention. In this embodiment, a camera 26, image acquisition means 31, individual information management means 32, planar dimension / area information acquisition means 33, side dimension / area information acquisition means 34, weight information acquisition means 40, and calibration unit 60 are provided. These means can be implemented by the CPU 10 appropriately reading a program from an external storage device 23 into the main memory 11, or by using a program that is initially stored in the main memory 11. Furthermore, the calibration unit 60 may be implemented by the CPU 10 appropriately reading a program from an external storage device 23 into the main memory 11, but it may also be implemented by a cloud computer connected to the computer in Figure 1 via a network. The calibration unit 60 is also provided in the weight measurement system according to the second embodiment of the present invention, and will be described in detail after the description of the weight measurement system according to the second embodiment of the present invention.

[0019] Camera 26 can be a 3D camera. Image acquisition means 31 captures images of the target animal using camera 26 and obtains a planar image from above and a lateral image from the side of the target animal. Individual information management means 32 uses the images obtained by image acquisition means 31 to perform individual identification of the target animal, issues and attaches identification information to the image of each individual target animal, and provides the image with this identification information to the planar dimension / area information acquisition means 33, the lateral dimension / area information acquisition means 34, and the weight information acquisition means 40.

[0020] The planar dimension / area information acquisition means 33 obtains planar dimension / area information of the target animal based on the planar image obtained by the image acquisition means. In this embodiment, planar dimension / area information refers to planar width information and planar length information, but of course, dimension information in diagonal directions on the plane or area information of a predetermined part of the plane may also be used. The planar width information of the target animal will be explained. Assuming that the image of the target animal from above (here, an image facing left) is shown as U in Figure 3, the width of the animal image (length in the vertical direction in the figure) is measured from the center to the left in the area obtained by dividing this image vertically in the figure, and the measurement value of the largest divided area is taken as the shoulder width (Figure 3(B)). Similarly, the measurement is taken from the center to the right, and the measurement value of the largest divided area is taken as the hip width (Figure 3(B)). These can be used as planar width information. In addition, hip width information, etc., defined by a predetermined definition may also be used as planar width information. The planar width information can be at least one of the target animal's shoulder width information, waist width information, or rump width information.

[0021] Next, we will explain the planar length information. As mentioned above, the positions of the shoulders and hips are determined in the process of obtaining the planar width information, so the distance between the shoulders and hips is taken as the body length. This body length can be used as the planar length information. The distance between the shoulders and waist can also be taken as the body length.

[0022] The side dimension / width information acquisition means 34 obtains the side dimension / width information of the target animal based on the side image obtained by the image acquisition means 31. In this embodiment, the side dimension / width information is used as body height information, but of course, dimension information such as diagonal dimensions on the side or width information of a predetermined part of the side may also be used. Side dimensions / width informationThis can be body height information, which is body height information for at least one of the shoulder, waist, and rump positions of the target animal. For example, as shown in Figure 4, using a planar image U and a lateral image S of the target animal, first, the highest point in the same divided area of ​​the lateral image S is determined in the length direction. Measurements are taken from the center of the image toward the left, and the height measurement of the divided area with the maximum is taken as the shoulder height. Measurements are taken from the center of the image toward the right, and the height measurement of the divided area with the maximum is taken as the rump height. Alternatively, the lateral dimensions and width information of the waist, etc., defined by a predetermined definition, may be used as planar width information. Side dimensions / width information This can be height information, which is height information of at least one of the shoulder, waist, and rump positions of the target animal.

[0023] The weight information acquisition means 40 includes a machine learning model 41 obtained by machine learning, with the planar dimensions / area information and the side dimensions / area information as explanatory variables and the weight information of the target animal as the target variable. This machine learning model 41 can be created by performing machine learning with the planar dimensions / area information and the side dimensions / area information obtained in this embodiment as explanatory variables and the weight information of the target animal measured as the target variable. The weight information acquisition means 40 provides the obtained planar dimensions / area information and the side dimensions / area information as explanatory variables to the machine learning model 41 to obtain the weight information, which is the target variable.

[0024] The weight measurement system with the above configuration is processed by a program corresponding to the flowchart shown in Figure 5. The operation will be explained below based on the flowchart. The target animal is imaged by camera 26 (S11). The process of imaged by camera 26 is continued in the subsequent steps.

[0025] When the camera 26 captures an image of the target animal, the CPU 10 receives the captured image data (S12) and calculates planar dimensions and width information, and lateral dimensions and width information (S13). The CPU 10 then provides the calculated planar dimensions and width information, and lateral dimensions and width information to a machine learning model to obtain the weight of the target animal (S14). In this way, it becomes possible to accurately measure the weight of the target animal.

[0026] Next, a weight measurement system according to a second embodiment will be described. Figure 6 shows a functional block diagram of the weight measurement system according to the second embodiment of the present invention. In this embodiment, the camera 26 and image acquisition means 31 are the same as in the first embodiment. In this embodiment, an individual information management means 32, an individual target animal image acquisition means 35, an appropriate image selection means 37, and a correction processing means 38 are provided. In this embodiment, the target animal is described as a pig, but the target animal of the present invention is not limited to this.

[0027] The individual information management means 32 uses the images obtained by the image acquisition means 31 to perform individual identification (image recognition) of the target animals, attaches identification information to the image of each individual target animal, and provides the images with this identification information to the planar dimension / area information acquisition means 33, the side dimension / area information acquisition means 34, and the weight information acquisition means 40. In other words, when multiple target animals are being raised, identification information is attached to each target animal to manage their images. More specifically, tracking is performed using RGB images, and a banding box is generated for each pig. If there are multiple pigs, a separate banding box and its coordinates are calculated for each pig, and these are passed to the individual target animal image acquisition means 35, which is the target selection unit.

[0028] This embodiment addresses cases where multiple target animals are kept together, and the camera 26 captures images of the area where multiple target animals are present. As a result, multiple target animals appear to overlap in the image. The individual target animal image acquisition means 35 separates the images of individual target animals from the images obtained by the camera 26 and acquires individual target animal images. In this embodiment, a segmentation method can be employed. By using this segmentation method, even overlapping pigs can be successfully separated, and mask images of each individual pig are generated and used for processing by the weight information acquisition means 40.

[0029] The appropriate image selection means 37 performs the process of selecting only images that can be properly processed when images with the identification information attached are provided to the planar dimension / area information acquisition means 33, the side dimension / area information acquisition means 34, and the weight information acquisition means 40. If the pig's posture is bent, the values ​​for each part cannot be measured correctly. Therefore, we considered what to do with images of the body in a bent state. In this embodiment, since a fixed-point camera is used, the same pig can be photographed multiple times, so we decided to discard the data of the body in a bent state and use only those photographed in a good state.

[0030] Therefore, we decided to exclude images where the back is curved. Next, the determination of whether the body is curved can be made as follows. When obtaining planar dimensions and area information, multiple divided areas are created as shown in Figure 3, and the line connecting the highest points in the height direction of each divided area in Figure 7 corresponds to the spine. The posture of the pig can be determined by detecting the degree of curvature of this spine. Since an ellipse has been fitted in the prior image processing, the center line of this ellipse is the ideal position of the spine. Images with a large deviation from this line are judged to have high curvature and can be excluded (Figure 7(B)), while those with a small deviation from this line can be kept (7(A)). For example, images where the distance from the line connecting the highest points in the height direction of each divided area is greater than a predetermined value can be excluded. Also, since even pigs judged to be normal have a slight curve, deviation data from the ideal position indicating this spinal curvature (e.g., variance) can also be added as an explanatory variable.

[0031] The correction processing means 38 performs a correction process on the image selected by the appropriate image selection means 37 to set a predetermined orientation. In this embodiment, the 3D image of one animal extracted by the individual target animal image acquisition means 35 using a segmentation method is used to measure each part using the planar dimension / area information acquisition means 33 and the side dimension / area information acquisition means 34. The parts to be measured can be shoulder width information, waist width information, hip width information, length from shoulder to hip, and the area of ​​the entire body. To measure each part, it is easier to calculate by slicing the image of the pig into cross-sections at regular intervals and processing them sequentially from the head, as described above. However, in this embodiment, since the pigs are photographed as they come to eat, images with various orientations are obtained. The posture is also not consistent, so ingenuity is required. The processing performed by the correction processing means 38, including solutions to these problems, will be described below.

[0032] <Correction process> If the image of the pig is cleanly cut out, a 3D image can be binarized and an ellipse approximation can be performed along the body's boundaries. The center of this ellipse will correspond to the center of the body, and the angle of the ellipse will directly match the angle of the pig. Therefore, we calculate the angle of the ellipse in the state shown in Figure 8(A), and rotate the image using an affine transformation with the center of the ellipse as the axis of rotation so that the angle of the ellipse that makes the image horizontal becomes horizontal (Figure 8(B)).

[0033] The weight measurement system with the above configuration processes data using a program corresponding to the flowchart shown in Figure 9. The operation will be explained below based on the flowchart. The CPU 10 captures an image of the target animal using the camera 26 (S11). Once the camera 26 has captured an image of the target animal, the CPU 10 receives the captured image data (S12) and separates the overlapping images of the target animals into a single image (S21).

[0034] Next, the CPU 10 performs individual identification (image recognition) and assigns identification information to each image of the target animal (S22). Then, it discards images that are distorted and selects only images that are captured in good condition (S23). The CPU 10 also performs elliptic approximation along the body boundary of the target animal, calculates the angle of the ellipse, and corrects the image by rotating it using an affine transformation with the center of the ellipse as the axis of rotation so that the angle of the ellipse that makes the image horizontal becomes horizontal (S24). Furthermore, it obtains planar dimensions and width information, and lateral dimensions and width information (S13). The CPU 10 provides the obtained planar dimensions and width information, and lateral dimensions and width information to a machine learning model to obtain the weight of the target animal (S14). In this way, it becomes possible to appropriately measure the weight of the target animal.

[0035] Furthermore, the weight information acquisition means 40 may calculate the average value of the weight information obtained within a predetermined time for a target animal identified as a single individual, and use the calculated average value as the weight of that target animal. Also, even if the target animal remains within the field of view, there may be cases where weight estimation is not possible due to poor conditions such as posture, which may result in fewer weight measurements being taken. If the acquired weight measurement value is below a set threshold, the acquired data may be discarded. Alternatively, if weight measurements are obtained more than the threshold number of times, the average may be calculated and the measurement may be completed as the weight measurement value for one target animal.

[0036] Furthermore, the weight measurement system may be equipped with a program that implements a species determination means (not shown) that determines species information within the target animal based on the planar dimensions / area information and the side dimensions / area information. In this case, the individual information management means 35 performs the operation of providing the planar dimensions / area information acquisition means 33, the side dimensions / area information acquisition means 34, and the weight information acquisition means 40 with an image that has identification information attached along with the species information determined by the species determination means, and the weight information acquisition means 40 may adjust the weight information using parameters corresponding to the species information within the target animal when obtaining weight information.

[0037] Each of the above embodiments can reduce the effort required to measure weight, enable daily management and data collection of pig weights, and lead to maximizing the profits of pig farms by optimizing the timing of shipment.

[0038] The configurations described above in the weight measurement system of the first embodiment and the weight measurement system of the second embodiment raise the following technical challenges: Specifically, are the accuracy of these embodiments sufficient for actual operation? Furthermore, can the same accuracy be ensured when deploying the system of this embodiment to various farms? And, can the system be operated with the same accuracy even if the breed of pig changes? In light of these challenges, the accuracy of the machine learning model 41 at a certain point in time was estimated and is shown in Figure 13. Specifically, the mean squared error (RMSE) was 2.6%. From this perspective, the weight measurement system of the first embodiment and the weight measurement system of the second embodiment of the present invention are equipped with a calibration unit 60. The calibration unit 60 will be described below.

[0039] The calibration unit 60 is equipped with a history information storage means 61, an updated learning model generation means 62, and an update timing control means 63. The history information storage means 61 stores pairs of explanatory variables and target variables as history information when weight information is acquired using the machine learning model 41 in the weight information acquisition means 40. The pairs of explanatory variables and target variables are stored in association with identification information issued for each individual by the individual information management means 32.

[0040] The updated learning model generation means 62 obtains an updated machine learning model by inputting the actual weight information of the target animal corresponding to the explanatory variables in the pairs of historical information stored in the historical information storage means 61, using this actual weight information as an explanatory variable and the actual weight information as the target variable, and performing the same machine learning as when the machine learning model 41 was created. More specifically, the updated machine learning model may be obtained when a predetermined number of the actual weight information to be used as the target variable is available. In other words, for example, the weight of the target animal can be measured using a scale at the time of shipment, and the data can be input to the calibration unit 60 in association with the identification information to create pairs of target variables and explanatory variables to be used in generating the updated machine learning model. Alternatively, it is also possible to use identification information from a meat processing plant and weight information obtained by measuring the carcass, which is the meat obtained by processing one individual animal.

[0041] Furthermore, if a sufficient number of target-and-explanatory variable pairs for generating the updated machine learning model cannot be obtained, the target-and-explanatory variable pairs created from the actual weight information may be mixed with the target-and-explanatory variable pairs used to generate the machine learning model 41. Also, if the weight information acquisition means 40 is used to adjust parameters according to the species information within the target animals when obtaining weight information, calibration may be performed to change these parameters according to the actual weight information. Moreover, calibration may be performed on the machine learning model 41 of the farm being calibrated using target-and-explanatory variable pairs created from actual weight information obtained from another farm with similar rearing methods and pig species ratios.

[0042] An update timing control means 63 is provided to determine the error between the weight information obtained using the machine learning model and the actual weight information. The update timing control means 63 creates error information between the weight information obtained using the machine learning model 41 and the actual weight information, and controls the timing of acquiring the updated machine learning model based on this error information. Therefore, the update timing control means 63 can be used to create and output error information between the weight information obtained using the machine learning model 41 and the actual weight information, and to monitor the error.

[0043] The update timing control means 63 can signal that it is time for an update when the mean squared error (RMSE) shown in Figure 13 exceeds, for example, 3%. Alternatively, the update learning model generation means 62 may be made to operate using pairs of target and explanatory variables created from the collected physical weight information, and the generated update learning model may be used as the new machine learning model 41.

[0044] According to the calibration unit 60, it is expected that the accuracy of this embodiment will be sufficient for actual operation, that the same accuracy can be ensured when deploying the system of this embodiment to various farms, and that the same accuracy can be maintained even if the breed of pig changes. [Explanation of Symbols]

[0045] 10 CPU 11 Main memory 12 buses 13 External Storage Interface 14 Input Interfaces 15 Display Interface 16. Data Input Interface 22 Pointing devices 23 External storage device 24 Input devices 25 Display device 26 cameras 31 Image acquisition method 32 Individual information management means 33. Means for acquiring planar dimensions and area information 34. Means for acquiring side dimensions and width information 35. Means for acquiring images of individual target animals 37. Appropriate Image Selection Method 38 Correction processing means 40. Means for acquiring weight information 41 Machine Learning Models 60 Calibration section 61. Means for storing historical information 62 Update learning model generation means 63 Update timing control means

Claims

1. Image acquisition means that captures images of the target animal using a camera and obtains a planar image from above and a lateral image from the side of the target animal, A means for acquiring planar dimensions and area information that obtains planar width information and planar length information of the target animal based on the aforementioned planar image, A means for obtaining lateral dimensions and width information of the target animal based on the aforementioned lateral image, The system includes a machine learning model obtained by machine learning with the aforementioned planar dimensions / area information and the aforementioned side dimensions / area information as explanatory variables and the weight information of the target animal as the objective variable, and a weight information acquisition means that provides the obtained planar dimensions / area information and the aforementioned side dimensions / area information as explanatory variables to the machine learning model to obtain the weight information, which is the objective variable, A calibration unit that performs calibration of the machine learning model, Individual information management means that uses the image obtained by the image acquisition means to perform individual identification of the target animal, attach identification information to the image of each individual target animal, and provide the image with the attached identification information to the planar dimension / area information acquisition means, the side dimension / area information acquisition means, and the weight information acquisition means. When providing the image with the aforementioned identification information to the planar dimension / area information acquisition means, the side dimension / area information acquisition means, and the weight information acquisition means, the appropriate image selection means determines the posture of the target animal by detecting the curvature of the spine, and selects only the images that can be processed appropriately based on this determination result. A weight measurement system comprising: a correction processing means that performs an elliptical approximation along the interface of the target animal's body on a planar image selected by the appropriate image selection means, calculates the angle of the ellipse, and performs a correction processing that rotates the image using an affine transformation with the center of the ellipse as the axis of rotation so that the angle of the ellipse that makes the image horizontal becomes horizontal; The system further includes a deviation data acquisition means that performs elliptical approximation along the body surface of the target animal to the planar image selected by the appropriate image selection means, and obtains deviation data between the length of the center line of the obtained ellipse and the length of the line segment obtained by connecting the highest points in each divided area obtained by dividing the planar image vertically into multiple sections. The aforementioned machine learning model is obtained by machine learning with the aforementioned planar dimensions / area information and side dimensions / area information as explanatory variables, the deviation data obtained by the deviation data acquisition means as explanatory variables, and the weight information of the target animal as the objective variable. The weight measurement system is characterized in that the weight information acquisition means adds the deviation data obtained by the deviation data acquisition means as an explanatory variable to the explanatory variables, namely the planar dimension / width information and the side dimension / width information, and provides this to the machine learning model to obtain the target variable, namely the weight information.

2. The calibration unit is The aforementioned weight information acquisition means includes a history information storage means that stores pairs of explanatory variables and target variables as history information when weight information is acquired using the machine learning model, Based on the input of the actual weight information of the target animal corresponding to the explanatory variables in the pairs of historical information stored in the historical information storage means, an updated learning model generation means obtains an updated machine learning model by performing the same machine learning as when the machine learning model was created, using the actual weight information as an explanatory variable and the actual weight information as an objective variable. The weight measurement system according to claim 1, characterized by comprising the following:

3. The weight measurement system according to claim 2, characterized in that the actual weight information is obtained by measuring the weight of the target animal using a weighing scale.

4. The weight measurement system according to claim 2, characterized in that the actual weight information is weight information obtained by measuring the weight of meat obtained by processing meat from the target animal.

5. The calibration unit includes: The weight measurement system according to claim 2, characterized in that it is provided with update timing control means that creates error information between weight information obtained using the machine learning model and actual weight information, and controls the timing of acquiring an updated machine learning model based on this error information.

6. The aforementioned camera captures images of the location where multiple target animals are present. The weight measurement system according to claim 1, further comprising individual target animal image acquisition means for separating images of individual target animals from images obtained by imaging with this camera and acquiring individual target animal images.

7. The weight measurement system according to claim 1, characterized in that the planar width information of the target animal is at least one of the shoulder width information of the target animal, the waist width information of the target animal, and the hip width information of the target animal.

8. The weight measurement system according to claim 1, characterized in that the planar length information of the target animal is the body length, and is the length from the shoulder to the rump of the target animal.

9. The weight measurement system according to claim 1, characterized in that the lateral dimensions and width information of the target animal is height information of at least one position of the target animal, such as the shoulder position, waist position, or hip position.

10. The weight measurement system according to claim 1, characterized in that the weight information acquisition means calculates the average value of weight information obtained within a predetermined time for a target animal identified as a single individual, and uses the calculated average value as the weight of the target animal.

11. The system includes a species determination means for determining species information within the target animal based on the aforementioned planar dimensions and area information and the aforementioned side dimensions and area information. The individual information management means performs the operation of providing the type information determined by the type determination means along with an image with the type information attached to it to the planar dimension / area information acquisition means, the side dimension / area information acquisition means, and the weight information acquisition means. The weight measurement system according to claim 1, characterized in that the weight information acquisition means adjusts using parameters corresponding to the type information when obtaining weight information.

12. Image acquisition step: The camera is used to image the target animal, and an image acquisition step is taken to obtain a planar image from above and a lateral image from the side of the target animal. A step to obtain planar dimensions and area information, in which planar dimensions and area information are obtained as planar width information and planar length information of the target animal based on the aforementioned planar image, A step of obtaining side dimensions and width information to obtain side dimensions and width information of the target animal based on the aforementioned side image, A weight information acquisition step involves using a machine learning model obtained by machine learning, with the aforementioned planar dimensions / area information and the aforementioned side dimensions / area information as explanatory variables and the weight information of the target animal as the objective variable, and providing the obtained planar dimensions / area information and the aforementioned side dimensions / area information as explanatory variables to the machine learning model to obtain the weight information, which is the objective variable. A historical information storage step is performed to store pairs of explanatory variables and target variables as historical information when weight information is acquired using the machine learning model in the weight information acquisition step, Based on the input of the actual weight information of the target animal corresponding to the explanatory variables in the pairs of historical information accumulated in the historical information accumulation step, an updated learning model generation step is performed to obtain an updated machine learning model by using this actual weight information as an explanatory variable and the actual weight information as the target variable, and performing the same machine learning as when the machine learning model was created. An individual information management step is performed which involves using the images obtained in the image acquisition step to perform individual identification of the target animals, attaching identification information to the image of each individual target animal, and providing the image with the attached identification information to the planar dimension / area information acquisition step, the side dimension / area information acquisition step, and the weight information acquisition step. When providing the image with the aforementioned identification information to the planar dimension / area information acquisition step, the side dimension / area information acquisition step, and the weight information acquisition step, the appropriate image selection step determines the posture of the target animal by detecting the curvature of the spine, and selects only the images that can be processed appropriately based on this determination result. A weight measurement method comprising: a correction processing step which involves applying an ellipse approximation along the interface of the target animal's body to the planar image selected in the appropriate image selection step, calculating the angle of the ellipse, and performing a correction processing to rotate the image to a predetermined orientation using an affine transformation with the center of the ellipse as the axis of rotation so that the angle of the ellipse that makes the image horizontal becomes horizontal; The system further includes a step to acquire deviation data, which involves performing an elliptical approximation along the body surface of the target animal with respect to the planar image selected in the appropriate image selection step, and obtaining deviation data between the length of the center line of the resulting ellipse and the length of the line segment obtained by connecting the highest points in each divided area obtained by dividing the planar image vertically into multiple sections. The aforementioned machine learning model is obtained by machine learning with the planar dimensions / area information and the side dimensions / area information, which are the explanatory variables, and the deviation data obtained in the deviation data acquisition step as explanatory variables, and the weight information of the target animal as the objective variable. The weight measurement method is characterized in that the weight information acquisition step adds the deviation data obtained in the deviation data acquisition step as an explanatory variable to the explanatory variables, namely the planar dimension / width information and the side dimension / width information, and provides these to the machine learning model to obtain the weight information, which is the target variable.

13. The weight measurement method according to claim 12, characterized in that the actual weight information is obtained by measuring the weight of the target animal using a weighing scale.

14. The method for measuring body weight according to claim 12, characterized in that the actual body weight information is weight information obtained by measuring the meat obtained by processing meat from the target animal.

15. The weight measurement method according to claim 12, characterized in that it includes an update timing control step that creates error information between weight information obtained using the machine learning model and actual weight information, and controls the timing of obtaining an updated machine learning model based on this error information.

16. The aforementioned camera captures images of the location where multiple target animals are present. The weight measurement method according to claim 12, further comprising an individual target animal image acquisition step of separating images of individual target animals from images obtained by imaging with this camera and acquiring individual target animal images.

17. The weight measurement method according to claim 12, characterized in that the planar width information of the target animal is at least one of the shoulder width information of the target animal, the waist width information of the target animal, and the hip width information of the target animal.

18. The method for measuring body weight according to claim 12, characterized in that the planar length information of the target animal is the length from the shoulder to the rump of the target animal.

19. The weight measurement method according to claim 12, characterized in that the lateral dimensions and width information of the target animal is height information of at least one position of the target animal, such as the shoulder position, waist position, or hip position.

20. The weight measurement method according to claim 12, characterized in that the weight information acquisition step involves determining the average value of weight information obtained within a predetermined time for a target animal identified as a single individual, and using the determined average value as the weight of the target animal.

21. The system includes a species determination step in which species information within the target animal is determined based on the aforementioned planar dimensions and area information and the aforementioned side dimensions and area information. The individual information management step performs the operation of providing the type information determined in the type determination step along with an image with the type information attached to it to the planar dimension / area information acquisition step, the side dimension / area information acquisition step, and the weight information acquisition step. The weight measurement method according to claim 12, characterized in that the weight information acquisition step is performed by adjusting the weight information using parameters corresponding to the type information when obtaining weight information.

Images