JP2018169993A - Result prediction apparatus, result prediction method, and program - Google Patents

Abstract

An object of the present invention is to predict a day when an object such as an optimal crop harvest date reaches a final result. A plurality of multi-wavelength photographic data, which is a multi-wavelength image obtained by a spectrum camera 1, is photographed at intervals before harvesting, together with photographing date data, for a crop that is a prediction target of a final result such as a harvest amount. The result state label that includes data up to the harvest date when the crop that has been collected and previously taken the multi-wavelength photo data is actually harvested is associated with the multi-wavelength photo data of the crop, and the multi-wavelength photo data and the result state label Is used to train the analysis parameters for outputting the result status label in response to the input of the multi-wavelength photographic data, and the individual before the harvest of the same type of crop is captured by the spectrum camera 1. Analyzing multi-wavelength photographic data, which is a wavelength image, using the above analysis parameters, and obtaining the corresponding result state label[Selection] Figure 1

Description

  The present invention creates a predictor using a large number of learning teacher data and AI technology when trying to predict temporal changes in an object to be photographed using an image in which multiple wavelengths are recorded. The present invention relates to an image analysis technique that can perform high time change prediction.

  Buildings installed outdoors and crops cultivated outdoors, unlike production activities in the factory, cannot simply achieve results as planned. Unlike industrial products, even when trying to improve the quality of crops uniformly, various factors such as temperature, humidity, sunshine hours, amount of water and fertilizer are involved. The environment is not completely controlled. For this reason, it continues to search for a method with high yield and certainty of how to grow and when to harvest.

  For example, since the growth process of tomatoes is complicated, it is difficult to accurately grasp the optimal harvest date. As a method for predicting the optimum harvest date, it has been conventionally performed based on numerical values such as an integrated temperature and an integrated amount of sunlight. However, the information alone was insufficient, and analysis with sufficient accuracy was not possible.

  Patent Document 1 proposes a method for performing irrigation control based on sensor data, a growth measuring device by aerial photography from a helicopter, and data of a monitoring device for observing individual crops equipped with a multispectral camera. Yes. A spectrum camera is a camera that can obtain more information than a conventional RBG data-only camera by measuring the color of a growing crop mainly over many frequency ranges. Specifically, the observation using a multispectral camera is performed with and without water stress, and an index PRI (Photochemical / Physiological reflectance Index) correlated with the photosynthetic rate is measured. It is described that irrigation control is performed to control water supply based on this.

  When analyzing multi-wavelength images taken with a spectral camera, it is generally performed by extracting a specific part of the image, creating a graph of wavelength intensity from the part, and analyzing the graph over time. It has been broken. The amount of information that can be obtained is particularly large for vegetables and fruits that change color at a timing suitable for harvest, such as tomatoes, apples, and strawberries.

  Further, Patent Document 2 proposes a method for predicting crop yield by combining images of survey areas and field attribute information accumulated in the past, aerial images of farm fields, and weather data. Yes.

  Patent Document 3 proposes a prediction program that predicts the amount of agricultural products that satisfy the shipping standards from the amount of delivered agricultural products.

JP2016-49102A JP-A-2015-49 Japanese Unexamined Patent Publication No. 2016-118895

  However, it is possible to observe the effect of irrigation control by analyzing multi-wavelength images taken with a spectrum camera that has more information than the conventional camera, but on the time axis where the final result can be obtained. There was not enough information to predict the future harvest date. Even if a multi-wavelength image of a growing agricultural product is taken, there is no uniform method for deriving the harvest date from that data. To do this, researchers had to have high skills and experience. For this reason, in practice, it is necessary to analyze the multi-wavelength images one by one, and it is difficult to increase the number of images to ensure accuracy.

  In addition, component values of multi-wavelength images are often recorded as wavelength combinations, but it is unknown which component (wavelength combination) the harvest date relates to, so examine all wavelength combinations. In order to increase information as a multi-wavelength image, the number of combinations increases explosively as the number of wavelengths increases, making analysis difficult.

  Accordingly, an object of the present invention is to enable prediction of future time points such as the date of harvest of agricultural products and the date of damage to infrastructure using a multi-wavelength image obtained by a spectrum camera.

This invention
Multi-wavelength photographic data, which is a multi-wavelength image obtained by a spectrum camera, is collected for the target object that is the result prediction target, and is collected together with the shooting date data by shooting multiple images at intervals before the final result.
Associating a result state label including data up to the final result date of the target object obtained by photographing the multi-wavelength photograph data in advance with the multi-wavelength photograph data of the target object,
Training analysis parameters for outputting the result state label with respect to the input of the multi-wavelength photograph data by the teacher data group that associates the multi-wavelength photograph data and the result state label,
Analyzing multi-wavelength photographic data, which is a multi-wavelength image photographed by a spectrum camera, for the sample before the final result of the same type of object using the analysis parameters to obtain the corresponding result state label, It was solved.

  For example, when the target object of the result prediction is a crop, the final result is a harvest, and the final result date is a harvest date. If the result prediction target is civil engineering and infrastructure such as bridges, overpasses, roads, and tunnels, the final result is damage such as cracking or peeling, and the final result date is the date of occurrence of the damage. The result prediction object handled in this invention is that the figure and state gradually change over time, but the change scale is several days or more, sometimes in units of years, so even with the human eye It is good that the time to reach the final result is hard to predict clearly.

  In the case of targeting crops, as a preparation stage, the daily state of the crops to be harvested is recorded as a multi-wavelength image, and an optimum harvest date is determined and harvested based on judgment by a skilled person. Thereby, it is possible to obtain a data group indicating how many days before the optimum harvest date each multi-wavelength image that has been photographed is recorded. The data up to the harvest date includes data indicating how many days before the actual harvest date. When the multi-wavelength photographic data of an unharvested crop is input using the analysis parameters trained by the data set associated with these, the number of days until the optimal harvest date can be obtained as the result state label as an output. This makes it possible not only to determine the optimal harvest date for each crop afterwards without the need of skilled experience, but also makes it possible to predict harvest based on criteria that cannot be judged by the human eye. .

  Further, the result status label may include not only the number of days until the final result date but also data regarding the quality and mass of the harvested crop such as sugar content and grade in the case of a crop. If the infrastructure is broken, data on the damage itself, such as an estimate of the work required for repair, may be included.

  For training of analysis parameters, a training method using supervised data in general machine learning can be used.

  The image by the spectrum camera can include the infrared region, and can obtain much more information than that seen by human eyes. By using analysis parameters trained with such a large amount of information, it is possible to predict the harvest date with higher accuracy.

  Even for vegetables and fruits that do not change in color before maturity, it is possible to capture changes in the size ratio of captured stems and branches and changes in the infrared region, etc. as criteria for judgment. It can be carried out. Moreover, even if the outdoor infrastructure is difficult for human eyes to understand because it is handled on a yearly scale, it can capture slight color changes on the surface, so that it can make predictions to a certain degree of accuracy.

Functional block diagram of a harvest prediction device and its peripheral devices according to an embodiment of the present invention FIG. 4 is a flowchart showing an example of a photographing step of the harvest prediction method according to the embodiment of the present invention. Flow diagram of training step of harvest prediction method according to one embodiment of this invention Flow diagram of prediction step of harvest prediction method according to one embodiment of this invention

  Hereinafter, the present invention will be described in detail together with specific embodiments. The present invention is an object result prediction method and result prediction apparatus for predicting a situation when an object whose state changes with time reaches some final result. First, as embodiments of a result prediction method and a result prediction apparatus, a crop yield prediction method and a harvest prediction apparatus capable of executing the same will be described.

  FIG. 1 shows a functional block diagram of a result prediction device 11 that is a harvest prediction device according to an embodiment of the present invention and its peripheral devices. The result prediction device 11 that performs prediction is connected to a data storage device 31 that stores image data to be used for training or to be predicted. The data storage device 31 may be the same housing as the result prediction device 11 or may be a separate housing. FIG. 1 shows a case of a separate housing. Moreover, it has the spectrum camera 1 which image | photographs a multiwavelength image. The spectrum camera 1 is connected to a photographing computer 2 that controls photographing. The photographing computer 2 is connected to the data storage device 31 by direct connection or network connection. FIG. 1 shows a case of connection via a LAN.

  The spectrum camera 1 is a camera having a larger recording wavelength than a general RGB three-primary color camera called a multispectral camera or a hyperspectral camera. It may be a collection of data at several tens of wavelengths, spectral data over a wide range in the visible light range, or spectral data over the infrared or ultraviolet range. Since the characteristics of crops tend to appear in the infrared region, it is preferable that images can be taken with a spectrum covering from the infrared region to the visible light region.

  The spectrum camera 1 captures the crop before harvesting together with the date data. As a teacher data group used for teacher data, which will be described later, a large number of data having different days until the actual final result date is required. This final result date is the harvest date in the case of crops. It is considered that there is a different feature between n days before the optimal harvest date and n + 1 days before, and it is necessary to train the analysis parameter 21 described later so that it can be observed with the spectrum camera 1 and found as a feature quantity. It is.

  The teacher data group is preferably as many as possible. For this reason, it is desirable not only to prepare a large number of data taken for separate crops of the same variety of crops, but also to take multiple shots at intervals for the same individual. As the shooting interval, it is desirable to take a photo every day or leave a longer interval. This is because the data for each day is sufficient because the date and time for harvesting are determined. However, in the case of crops whose final quality differs greatly in time units, images may be taken every 3 to 12 hours. On the other hand, if the shooting interval exceeds one week, the accuracy of the data up to the harvest date decreases, and it is difficult to increase the number of data. For this reason, the imaging interval for the same individual is preferably 7 days or less, and more preferably 4 days or less. Such imaging is performed for as many individuals as possible. However, the imaging interval may be different for each individual.

  On the other hand, with respect to an individual of a crop to be predicted for a harvest date or the like by the result predicting apparatus 11 according to the present invention, it is possible to perform the prediction according to the present invention even with data taken once.

  It is desirable that an image captured by the spectrum camera 1 basically includes an eating part such as a fruit of a crop. However, depending on the variety, the production of the crop is judged from the state of the leaves and stems. It is preferable to photograph the leaves of root vegetables and potatoes that eat in the soil. On the other hand, in the case of flowers, it is necessary to include a bud portion.

  The photographing computer 2 is connected to the spectrum camera 1 and controls to periodically photograph the spectrum camera 1. It may be fixedly installed in front of the crop, may be moved along a rail installed on the farm, or may be moved in the air by a drone. In FIG. 1, the connection is a wired connection such as USB, but a wireless connection may be used if the spectrum camera 1 has a sufficient wireless connection function.

  The photographing computer 2 has camera operation software 3 for controlling the spectrum camera 1. When the photographing computer 2 is a personal computer, the camera operation software 3 displays an item for controlling the spectrum camera 1 on the screen, and accepts an operation with a pointing device, a keyboard, or the like. When the photographing computer 2 is a server, the camera operation software 3 further has a network server function for receiving operations from a personal computer, a smartphone, a tablet, or the like connected via a network. Regardless of the form, the camera operation software 3 is used to perform control so that a photograph of a crop to be photographed is regularly photographed at the photographing interval.

  The photographing computer 2 has a storage unit 4 that records multi-wavelength photo data, which is a multi-wavelength image photographed by the spectrum camera 1, together with date / time data of photographing. The hardware of the storage unit 4 may be a magnetic disk or a non-volatile semiconductor memory.

  The imaging computer 2 includes a captured image communication unit 5 that is a network interface that is directly or indirectly connected to the data storage device 31. Further, when the multi-wavelength photographic data is recorded in the storage unit 4, the captured image control unit 6 that controls the imaging computer 2 transfers the data to the data storage device 31 sequentially or batchwise.

  The data storage device 31 may be a server having a large-scale storage device. It is only necessary that images can be stored and read sequentially. Specifically, a commercially available NAS (Network Attached Storage) may be used, a personal computer may be used, or a dedicated data server may be used. It may also exist on the cloud. When the storage area communication unit 32 that controls the communication interface of the data storage device 31 receives the multi-wavelength photo data from the imaging computer 2, the storage region control unit 33 that controls the data storage device 31 receives the multi-wavelength photo data. Is stored in the storage device storage unit 34. In the storage device storage unit 34, the photographed photos are classified for each photographing object or classified for each photographing date and time so as to be easily identified. For example, as shown in FIG. 1, there is a mode in which a learning image storage unit 35 that collects data used for a teacher data group and a prediction image storage unit 36 that collects data for which a harvest date or the like is to be predicted are collected. Can be mentioned. The learning image storage unit 35 may further divide folders for each shooting date.

  The result prediction device 11 performs registration of the multi-wavelength photograph data stored in the data storage device 31 in association with the harvest date and training of the analysis parameters for the prediction program using the registration. Further, after the training, for the multi-wavelength photo data stored in the data storage device 31, an output that is a harvest date prediction or other harvest prediction is obtained.

  First, the components involved in the training will be described. The learning image control unit 12 of the result prediction apparatus 11 reads the multi-wavelength photographic data from the data storage device 31 via the analysis image communication unit 13 that controls the network interface.

  The state label input unit 16 of the result prediction apparatus 11 accepts the input of the actual harvest date of each of the captured crops for each of the read multi-wavelength photo data. In order to prompt the input, the colorization processing unit 15 converts the read multi-wavelength photo data into a color image that can be easily discerned by human eyes. This is because images over a wide range may be difficult to recognize by human eyes. Basically, it is preferable that the image is composed of three colors of red, green, and blue (RGB), but any image that can be recognized by a monochrome image can be used. Further, the color is not limited to simple red, green and blue, and a pseudo color may be obtained by appropriately correcting wavelengths including the infrared region and the ultraviolet region. In any form, the photograph of the displayed crop is seen so that the worker can recognize the actual harvest date.

  The state label input unit 16 of the result prediction apparatus 11 outputs the converted image to the monitor, and accepts the input of the harvest date. Based on the input harvest date and the shooting date of each multi-wavelength photo data, the multi-wavelength photo data is calculated backward from the harvest date to calculate how many days ago the data is. The days data until the harvest date (remaining days data) is associated with each multi-wavelength photo data as a result state label.

  Note that the content of the state label input unit 16 is not limited to the above-described content requesting input by human power. For example, if multi-wavelength photographic data is classified in advance for each house or per basket, if they are harvested together for each house or per basket, the same harvest date will be used, so input is performed in batch processing. be able to. In addition, even when the harvest date is different, for example, each crop is managed by RFID, and each crop is identified and photographed by RFID even when the multi-wavelength photo data is taken. The processed data may be collectively read and input as a result state label.

  Further, the result state label input by the state label input unit 16 is not limited to the remaining number of days data. Other information necessary as an agricultural product such as sugar content, moisture content, salt content, mass, evaluated grade, and shipping value measured after harvesting may be registered together.

  The state label adding unit 17 of the result prediction apparatus 11 associates the input result state label with the multi-wavelength photograph data. These associated data groups are used as a teacher data group. This data set may be sent to the next data preprocessing unit 18 as it is, or the data once associated may be sent to the data storage device 31 and recorded as a set. In that case, when training to be described later is performed at a later date, the multi-wavelength photograph data can be read together with the result status label, and training can be performed without inputting the result status label.

  The data pre-processing unit 18 of the result prediction apparatus 11 preferably performs pre-processing of the read multi-wavelength photograph data according to a pre-processing flow set in advance by the user, and uses the pre-process after performing this pre-processing. Before being read by the AI learning unit 19 to be described later, the white balance, saturation, brightness, whiteness, etc. of the photo are adjusted, and the above-mentioned analysis parameters are mainly different in the conditions at the time of photographing the data to be read. Try to minimize the impact.

  The AI learning unit 19 of the result prediction apparatus 11 reads the teacher data group that is the multi-wavelength photograph data associated with the result state label. At this time, the multi-wavelength photographic data is preferably pre-processed. Analyzing parameters between inputs and outputs of a neural network that operates artificial intelligence software as teacher data, which is obtained by subdividing the multi-wavelength photographic data and inputting to the artificial intelligence software 20 having a neural network and outputting the result state label. Train 21. The neural network preferably supports multi-layered so-called deep learning. This training is performed using as many teacher data as possible, and the analysis parameter 21 is adjusted. The result prediction apparatus 11 stores the analysis parameter 21 adjusted in this way, and uses it for harvest prediction described later.

  Next, components related to the harvest prediction which is the result prediction will be described. The prediction image control unit 22 of the result prediction apparatus 11 reads the multi-wavelength photo data from the data storage device 31 via the analysis image communication unit 13 that controls the network interface. The data to be read is the multi-wavelength photograph data that reflects the contents of the result state label in the test. What is used in the actual prediction is the multi-wavelength photographic data taken for the individual crops to be harvested. A large number of data may be read at once, or all the multi-wavelength photo data for which the harvest prediction is to be obtained may be read at a time.

  It is preferable that the data preprocessing unit 18 of the result prediction apparatus 11 preprocesses the multi-wavelength photographic data used for harvest prediction according to the same preprocessing flow as described above. Prediction accuracy is improved by adjusting the white balance, brightness, and the like of the data group to be read to the same conditions as the data group used for training.

  The AI prediction unit 23 of the result prediction apparatus 11 reads the multi-wavelength photographic data used for the harvest prediction, and inputs it to the artificial intelligence software 20 using the adjusted analysis parameter 21. In addition, the artificial intelligence software 20 may separately provide a command for input so that a function or condition value, which is a condition of the neural network, can be adjusted, and reflect the value. In response to the input, the artificial intelligence software 20 derives the contents of the result state label. The derived result state label has parameters having the same attributes as the result state label used for training. That is, in addition to the number of days until the harvest date, if the sugar content, water content, mass, etc. are used, those properties can also be derived as prediction data. The derived result state label is stored in the storage unit 25 of the result prediction apparatus 11.

  When the result state label derived in the storage unit 25 is stored, the prediction label output unit 24 of the result prediction apparatus 11 reads the result, shapes it, and outputs it. Specifically, the optimal harvest date and other harvest status predictions are displayed and printed on an attached monitor or printer.

  An embodiment example of a harvest prediction method using the result prediction apparatus 11 will be described. These flows are divided into three stages: a shooting step, a training step, and a prediction step.

  First, the above photographing step will be described with reference to the flow example of FIG. 2 (S101). The individual crops are photographed using the camera operation software 3, and the multi-wavelength photograph data is stored in the storage unit 4 such as an HDD together with the photographing date data (S102). The captured image control unit 6 monitors the storage of the multi-wavelength photo data in the storage unit 4 (S103), and when the file generation of the multi-wavelength photo data is confirmed, the data storage device 31 via the captured image communication unit 5. The multi-wavelength photo data is transferred to (S104).

  When the storage area control unit 33 of the data storage device 31 receives the multi-wavelength photo data via the storage area communication unit 32 (S111), the multi-wavelength photo to the learning image storage unit 35 and the prediction image storage unit 36 is stored. Data files are sorted (S112). The sorting conditions are not particularly limited, such as varieties, farmland, and shooting dates, and are set arbitrarily by the user. For example, for a single type of crop, a folder is created for each shooting date, and the multi-wavelength photo data is stored therein. The user may specify to the storage area control unit 33 whether the multi-wavelength photo data is allocated to the learning image storage unit 35 or the prediction image storage unit 36. Here, if the multi-wavelength photo data used for the teacher data group is stored in the learning image storage unit 35 (S113), if it is the multi-wavelength photo data for actual prediction, the prediction image storage is performed. Store in the unit 36 (S114).

  Next, the training step will be described using the flow example of FIG. First (S201), the result prediction apparatus 11 prepares the artificial intelligence software 20 having the initialized analysis parameter 21 (S202). The learning image control unit 12 transmits a transfer request for the teacher data group to the storage area control unit 33 via the analysis image communication unit 13, and receives the teacher data group from the learning image storage unit 35 (S203). ). At this time, all the multi-wavelength photo data stored in the learning image storage unit 35 may be requested in accordance with circumstances such as calculation resources of the result prediction apparatus 11 or a part of the multi-wavelength photo data. You may request only.

The colorization processing unit 15 converts the learning multi-wavelength image received by the learning image control unit 12 into a color image including three colors of red, green, and blue, and transfers the color image to the state label input unit 16 (S211).
In this conversion, a pseudo color image may be generated using wavelengths other than the above three colors, or a more detailed color image may be generated by assigning appropriate weights to all wavelengths. The state label input unit 16 displays a color image on a monitor (not shown) (S212), and inputs a result state label including harvest date data of individual crops shown in the color image by an input interface such as a keyboard. (S213), and the input is transmitted to the state label adding unit 17 (S214). Instead of S213, a list of result status labels prepared in advance may be read as input.

  When receiving the transmission of the result state label from the state label input unit 16, the state label adding unit 17 promptly requests the learning data group for learning to be associated with the learning image control unit 12 (S221). When receiving the request from the state label adding unit 17, the learning image control unit 12 immediately transfers the multi-wavelength photo data to the data preprocessing unit 18. The data preprocessing unit 18 performs processing for aligning the brightness and the like on the inputted multi-wavelength photo data according to a predetermined preprocessing flow (S222), and transfers the preprocessed teacher data group to the state label adding unit. The state label adding unit 17 adds the above result state label to the transferred preprocessed teacher data group (S223), and transfers the result state label to the AI learning unit 19.

  The AI learning unit 19 introduces the result state label associated with the received preprocessed teacher data group into the artificial intelligence software 20 as a set, and adjusts the analysis parameter 21 (S224). The multi-wavelength photo data is input as the teacher data group, and this is repeated until the output result state label converges (S225), and the finally obtained analysis parameter 21 is stored in the storage unit 25 (S226). ).

  And said prediction step is demonstrated using the example of a flow of FIG. 4 (S301). The prediction image control unit 22 transmits a transfer request for the prediction image to the storage area control unit 33 via the analysis image communication unit 13, receives the multi-wavelength photograph data from the prediction image storage unit 36 (S302), The AI prediction unit 23 is notified. What is required at this time is the multi-wavelength photographic data of the individual that is the target for which the result state label including the harvest date is to be predicted. In order to increase the prediction accuracy by taking the average value, a plurality of data may be read at a time, all the multi-wavelength photo data stored in the prediction image storage unit 36 may be requested, or some of the above Only multi-wavelength photo data may be requested.

  Based on the notification from the predicted image control unit 22, the AI prediction unit 23 requests the predicted image control unit 22 to perform preprocessing for using the multi-wavelength photograph data for prediction. When receiving the request from the AI prediction unit 23, the predicted image control unit 22 immediately transfers the multi-wavelength photo data to the data preprocessing unit 18. The data preprocessing unit 18 processes the input multi-wavelength photograph data under the same conditions as the preprocessing flow (S222) (S303), and transfers the preprocessed prediction image to the AI prediction unit 23. The AI prediction unit 23 outputs the received preprocessed prediction image to the artificial intelligence software 20. Using the input preprocessed prediction image and the analysis parameter 21, the artificial intelligence software derives the harvest prediction label (S304), and stores the output result of the harvest prediction label in the storage unit 25 of the result prediction apparatus 11. (S305). The prediction label output unit 24 monitors the storage unit 25. When the generation of the derived harvest prediction label is confirmed, the prediction label output unit 24 immediately reads the output result, shapes it, and outputs it to a monitor (not shown) (S306).

  Other embodiments of the result prediction method and result prediction apparatus according to the present invention include an infrastructure aging prediction method and an aging prediction apparatus capable of executing the same. The basic configuration is the same as that of the result prediction apparatus 11 represented in FIG. 1, and the object photographed by the spectrum camera 1 serves as an infrastructure for predicting aging. However, the aging time of the infrastructure is longer than the time until the crop is harvested, and the monthly or yearly photographing is periodically performed at intervals.

  The target of the infrastructure to be photographed is mainly the surface of the civil engineering infrastructure made of concrete. You can use tunnels, viaducts, buildings, etc. that are deteriorated by exposure to wind and rain outdoors. Changes in moisture contained in the concrete and elution of metal ions accompanying deterioration of the internal reinforced concrete are recognized as information by the spectrum camera 1 that images the concrete surface. In addition, the progress of rust on the surface of steel towers, bridges, and wires may be recognized as information by the spectrum camera 1. In any case, the end point of aging is the date of the failure of those infrastructures.

  As in the case of the crops embodiment, the data storage unit for the data used for the teacher data and the data storage unit for predicting the damage date corresponding to the final result date may be separated, but it takes a long time to detect the infrastructure damage. In particular, it is good to continue to shoot and accumulate data without limiting which one is used, and to use the data of an object that has actually been damaged as a teacher data group. When the number of data that can be used as the teacher data group is sufficient, the analysis parameters are adjusted by the training step, and then the target data that has not yet been damaged is output as the prediction of the damage date and other results.

DESCRIPTION OF SYMBOLS 1 Spectrum camera 2 Shooting computer 3 Camera operation software 4 Memory | storage part 5 Captured image communication part 6 Captured image control part 11 Result prediction apparatus 12 Learning image control part 13 Analytical image communication part 15 Colorization process part 16 State label input part 17 State Label adding unit 18 Data preprocessing unit 19 AI learning unit 20 Artificial intelligence software 21 Analysis parameter 22 Predictive image control unit 23 AI prediction unit 24 Predictive label output unit 25 Storage unit 31 Data storage device 32 Storage region communication unit 33 Storage region control unit 34 Storage Device Storage Unit 35 Learning Image Storage Unit 36 Prediction Image Storage Unit

Claims (4)

This is a collection of multi-wavelength photographic data that is a multi-wavelength image, including the shooting date data, taken by a spectrum camera for a sample group of the target object that is the result prediction target, and the remaining number of days until the final result date of the later date For a data group in which different data exists, a state label adding unit that obtains a teacher data group by associating a result state label determined after the final result including the remaining number of days until the final result date of the later date,
An AI learning section that reads the teacher data group, has a neural network that sets the multi-wavelength photograph data as input, and sets the result state label as output, and adjusts analysis parameters between input and output;
The multi-wavelength photograph data taken for the sample of the same kind as the object and before the final result is read into the neural network, and the result state label is derived as a predicted value according to the adjusted analysis parameter. A result prediction apparatus comprising: an AI prediction unit. Multi-wavelength photographic data, which is a multi-wavelength image obtained by a spectrum camera, is collected for the target object that is the result prediction target, and is collected together with the shooting date data by shooting multiple images at intervals before the final result.
Associating a result state label including data up to the final result date of the target object obtained by photographing the multi-wavelength photograph data in advance with the multi-wavelength photograph data of the target object,
Training analysis parameters for outputting the result state label with respect to the input of the multi-wavelength photograph data by the teacher data group that associates the multi-wavelength photograph data and the result state label,
A result prediction method for analyzing the multi-wavelength photographic data, which is a multi-wavelength image photographed by a spectrum camera, of the sample before the final result of the same type of object using the analysis parameter to obtain the corresponding result state label. Multi-wavelength photographic data that is a multi-wavelength image obtained by capturing the target object that is the result prediction target with a spectrum camera is used as input side data,
An analysis parameter for a neural network, in which a result state label including data up to a final result date at which the object obtained by photographing the multi-wavelength photo data reaches a final result is output side data. The multi-wavelength photographic data is input to a neural network having the analysis parameter according to claim 3,
A result prediction program that outputs the result status label.

Images